{
  "$comment": "GENERATED by scripts/ssg.ts from lib/benchmarks/dataset.ts. Do not hand-edit. The measurements this is built on live at /measured-benchmarks.json, which keeps its own stable shape.",
  "schemaVersion": "2.0",
  "generatedAt": "2026-08-17",
  "license": "CC BY 4.0",
  "attribution": "LLM Configurator — https://llmconfigurator.com/en/benchmarks",
  "provenanceNote": "Every figure carries its own provenance. \"measured\" means a published or submitted benchmark run for this exact combination. \"estimated\" means model output: \"fitted\" where the architecture's efficiency constant was solved from at least 3 measured runs, \"provisional\" where it was solved from fewer (too few to separate the model's free parameters), and \"uncalibrated\" where it was never solved at all. A cell whose model spills to system RAM is flagged: that path is calibrated against no measurements whatsoever.",
  "schemaDescription": "One record per (accelerator, model) pair at a fixed quantization and context. decodeTokS is decode throughput; decodeLow/High bound the reported interval. wattsBasis distinguishes discrete-card board power from unified-memory SoC package power — they are not the same envelope and must not be compared as though they were. provenance.speed and provenance.power are independent: a record may be measured on one and estimated on the other.",
  "parameters": {
    "quant": "Q4_K_M",
    "contextLength": 4096,
    "kwhRateUsd": 0.15,
    "note": "Cells are computed at one quant and context. The interactive database at /en/benchmarks covers the full matrix."
  },
  "calibration": {
    "mapePct": 7.68,
    "worstResidualPct": 31.53,
    "calibrationRuns": 14,
    "minRunsForConfidentFit": 3,
    "architectures": [
      {
        "architecture": "nvidia-blackwell",
        "state": "seed",
        "fittedFromRuns": 0,
        "eta": 0.85,
        "mapePct": null
      },
      {
        "architecture": "nvidia-ada",
        "state": "fitted",
        "fittedFromRuns": 12,
        "eta": 0.927,
        "mapePct": 7.68
      },
      {
        "architecture": "nvidia-ampere",
        "state": "provisional",
        "fittedFromRuns": 1,
        "eta": 0.6223,
        "mapePct": null
      },
      {
        "architecture": "nvidia-hopper",
        "state": "seed",
        "fittedFromRuns": 0,
        "eta": 0.85,
        "mapePct": null
      },
      {
        "architecture": "amd-rdna4",
        "state": "seed",
        "fittedFromRuns": 0,
        "eta": 0.7,
        "mapePct": null
      },
      {
        "architecture": "amd-rdna3",
        "state": "seed",
        "fittedFromRuns": 0,
        "eta": 0.65,
        "mapePct": null
      },
      {
        "architecture": "intel-xe2",
        "state": "seed",
        "fittedFromRuns": 0,
        "eta": 0.6,
        "mapePct": null
      },
      {
        "architecture": "apple-silicon",
        "state": "provisional",
        "fittedFromRuns": 1,
        "eta": 0.5481,
        "mapePct": null
      },
      {
        "architecture": "generic",
        "state": "seed",
        "fittedFromRuns": 0,
        "eta": 0.75,
        "mapePct": null
      },
      {
        "architecture": "cpu",
        "state": "seed",
        "fittedFromRuns": 0,
        "eta": 0.35,
        "mapePct": null
      }
    ],
    "offloadTerm": {
      "state": "uncalibrated",
      "note": "No published benchmark we could source runs with weights spilled to host RAM in a way we could reproduce — the three candidates found are listed with reasons in lib/data/efficiency-fit.json under excludedPoints. Any cell flagged offloadWarning rests on this term."
    }
  },
  "counts": {
    "cells": 446,
    "accelerators": 109,
    "measuredRuns": 14,
    "publishers": [
      "Hardware Corner",
      "Mustafa.net"
    ]
  },
  "cells": [
    {
      "gpuId": "rtx-5090-laptop",
      "gpuName": "NVIDIA GeForce RTX 5090 Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 107.61,
      "decodeLowTokS": 69.95,
      "decodeHighTokS": 145.27,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 150,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.72,
      "localCostPer1MUsd": 0.05808,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (150 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (150 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5090-laptop",
      "gpuName": "NVIDIA GeForce RTX 5090 Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 116.56,
      "decodeLowTokS": 75.76,
      "decodeHighTokS": 157.36,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 150,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.78,
      "localCostPer1MUsd": 0.05362,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (150 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (150 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5090-laptop",
      "gpuName": "NVIDIA GeForce RTX 5090 Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 67.3,
      "decodeLowTokS": 43.74,
      "decodeHighTokS": 90.85,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 150,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.45,
      "localCostPer1MUsd": 0.09287,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (150 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (150 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5090-laptop",
      "gpuName": "NVIDIA GeForce RTX 5090 Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 185.5,
      "decodeLowTokS": 120.58,
      "decodeHighTokS": 250.43,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 150,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 1.2,
      "localCostPer1MUsd": 0.03369,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (150 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (150 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5090-laptop",
      "gpuName": "NVIDIA GeForce RTX 5090 Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 104.83,
      "decodeLowTokS": 68.14,
      "decodeHighTokS": 141.52,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 150,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.7,
      "localCostPer1MUsd": 0.05962,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (150 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (150 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m5-ultra",
      "gpuName": "Apple M5 Ultra",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 91.18,
      "decodeLowTokS": 59.27,
      "decodeHighTokS": 123.09,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 120,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.76,
      "localCostPer1MUsd": 0.05484,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (120 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (120 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m5-ultra",
      "gpuName": "Apple M5 Ultra",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 98.42,
      "decodeLowTokS": 63.97,
      "decodeHighTokS": 132.87,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 120,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.82,
      "localCostPer1MUsd": 0.0508,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (120 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (120 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m5-ultra",
      "gpuName": "Apple M5 Ultra",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 57.93,
      "decodeLowTokS": 37.66,
      "decodeHighTokS": 78.21,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 120,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.48,
      "localCostPer1MUsd": 0.08631,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (120 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (120 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m5-ultra",
      "gpuName": "Apple M5 Ultra",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 152.54,
      "decodeLowTokS": 99.15,
      "decodeHighTokS": 205.93,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 120,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.3,
      "localCostPer1MUsd": 0.03278,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (120 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (120 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m5-ultra",
      "gpuName": "Apple M5 Ultra",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 88.92,
      "decodeLowTokS": 57.8,
      "decodeHighTokS": 120.04,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 120,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.74,
      "localCostPer1MUsd": 0.05623,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (120 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (120 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "arc-pro-b70",
      "gpuName": "Intel Arc Pro B70",
      "vendor": "Intel",
      "architecture": "intel-xe2",
      "backend": "vulkan",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 53.44,
      "decodeLowTokS": 34.74,
      "decodeHighTokS": 72.15,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 225,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.24,
      "localCostPer1MUsd": 0.17542,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "intel-xe2",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (225 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (225 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "arc-pro-b70",
      "gpuName": "Intel Arc Pro B70",
      "vendor": "Intel",
      "architecture": "intel-xe2",
      "backend": "vulkan",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 58.07,
      "decodeLowTokS": 37.74,
      "decodeHighTokS": 78.39,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 225,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.26,
      "localCostPer1MUsd": 0.16146,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "intel-xe2",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (225 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (225 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "arc-pro-b70",
      "gpuName": "Intel Arc Pro B70",
      "vendor": "Intel",
      "architecture": "intel-xe2",
      "backend": "vulkan",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 32.97,
      "decodeLowTokS": 21.43,
      "decodeHighTokS": 44.5,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 225,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.15,
      "localCostPer1MUsd": 0.28438,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "intel-xe2",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (225 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (225 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "arc-pro-b70",
      "gpuName": "Intel Arc Pro B70",
      "vendor": "Intel",
      "architecture": "intel-xe2",
      "backend": "vulkan",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 94.65,
      "decodeLowTokS": 61.52,
      "decodeHighTokS": 127.78,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 225,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.42,
      "localCostPer1MUsd": 0.09905,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "intel-xe2",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (225 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (225 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "arc-pro-b70",
      "gpuName": "Intel Arc Pro B70",
      "vendor": "Intel",
      "architecture": "intel-xe2",
      "backend": "vulkan",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 52.01,
      "decodeLowTokS": 33.81,
      "decodeHighTokS": 70.21,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 225,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.23,
      "localCostPer1MUsd": 0.18025,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "intel-xe2",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (225 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (225 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "arc-pro-b65",
      "gpuName": "Intel Arc Pro B65",
      "vendor": "Intel",
      "architecture": "intel-xe2",
      "backend": "vulkan",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 53.44,
      "decodeLowTokS": 34.74,
      "decodeHighTokS": 72.15,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 200,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.27,
      "localCostPer1MUsd": 0.15593,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "intel-xe2",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (200 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (200 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "arc-pro-b65",
      "gpuName": "Intel Arc Pro B65",
      "vendor": "Intel",
      "architecture": "intel-xe2",
      "backend": "vulkan",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 58.07,
      "decodeLowTokS": 37.74,
      "decodeHighTokS": 78.39,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 200,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.29,
      "localCostPer1MUsd": 0.14352,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "intel-xe2",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (200 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (200 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "arc-pro-b65",
      "gpuName": "Intel Arc Pro B65",
      "vendor": "Intel",
      "architecture": "intel-xe2",
      "backend": "vulkan",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 32.97,
      "decodeLowTokS": 21.43,
      "decodeHighTokS": 44.5,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 200,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.16,
      "localCostPer1MUsd": 0.25278,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "intel-xe2",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (200 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (200 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "arc-pro-b65",
      "gpuName": "Intel Arc Pro B65",
      "vendor": "Intel",
      "architecture": "intel-xe2",
      "backend": "vulkan",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 94.65,
      "decodeLowTokS": 61.52,
      "decodeHighTokS": 127.78,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 200,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.47,
      "localCostPer1MUsd": 0.08804,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "intel-xe2",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (200 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (200 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "arc-pro-b65",
      "gpuName": "Intel Arc Pro B65",
      "vendor": "Intel",
      "architecture": "intel-xe2",
      "backend": "vulkan",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 52.01,
      "decodeLowTokS": 33.81,
      "decodeHighTokS": 70.21,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 200,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.26,
      "localCostPer1MUsd": 0.16022,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "intel-xe2",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (200 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (200 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "radeon-ai-pro-r9700",
      "gpuName": "AMD Radeon AI PRO R9700",
      "vendor": "AMD",
      "architecture": "amd-rdna4",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 64.6,
      "decodeLowTokS": 41.99,
      "decodeHighTokS": 87.21,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.22,
      "localCostPer1MUsd": 0.1935,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna4",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "radeon-ai-pro-r9700",
      "gpuName": "AMD Radeon AI PRO R9700",
      "vendor": "AMD",
      "architecture": "amd-rdna4",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 70.09,
      "decodeLowTokS": 45.56,
      "decodeHighTokS": 94.63,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.23,
      "localCostPer1MUsd": 0.17833,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna4",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "radeon-ai-pro-r9700",
      "gpuName": "AMD Radeon AI PRO R9700",
      "vendor": "AMD",
      "architecture": "amd-rdna4",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 40.09,
      "decodeLowTokS": 26.06,
      "decodeHighTokS": 54.12,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.13,
      "localCostPer1MUsd": 0.31179,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna4",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "radeon-ai-pro-r9700",
      "gpuName": "AMD Radeon AI PRO R9700",
      "vendor": "AMD",
      "architecture": "amd-rdna4",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 113.04,
      "decodeLowTokS": 73.48,
      "decodeHighTokS": 152.61,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.38,
      "localCostPer1MUsd": 0.11058,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna4",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "radeon-ai-pro-r9700",
      "gpuName": "AMD Radeon AI PRO R9700",
      "vendor": "AMD",
      "architecture": "amd-rdna4",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 62.9,
      "decodeLowTokS": 40.88,
      "decodeHighTokS": 84.91,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.21,
      "localCostPer1MUsd": 0.19874,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna4",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3090-ti",
      "gpuName": "NVIDIA GeForce RTX 3090 Ti",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 92.58,
      "decodeLowTokS": 63.39,
      "decodeHighTokS": 121.77,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 450,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.21,
      "localCostPer1MUsd": 0.20253,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (450 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (450 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3090-ti",
      "gpuName": "NVIDIA GeForce RTX 3090 Ti",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 100.65,
      "decodeLowTokS": 68.92,
      "decodeHighTokS": 132.39,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 450,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.22,
      "localCostPer1MUsd": 0.18629,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (450 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (450 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3090-ti",
      "gpuName": "NVIDIA GeForce RTX 3090 Ti",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 56.95,
      "decodeLowTokS": 38.99,
      "decodeHighTokS": 74.9,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 450,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.13,
      "localCostPer1MUsd": 0.32926,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (450 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (450 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3090-ti",
      "gpuName": "NVIDIA GeForce RTX 3090 Ti",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 164.9,
      "decodeLowTokS": 112.91,
      "decodeHighTokS": 216.9,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 450,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.37,
      "localCostPer1MUsd": 0.1137,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (450 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (450 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3090-ti",
      "gpuName": "NVIDIA GeForce RTX 3090 Ti",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 90.08,
      "decodeLowTokS": 61.68,
      "decodeHighTokS": 118.48,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 450,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.2,
      "localCostPer1MUsd": 0.20815,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (450 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (450 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3080-ti",
      "gpuName": "NVIDIA GeForce RTX 3080 Ti",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 85.45,
      "decodeLowTokS": 58.51,
      "decodeHighTokS": 112.4,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 350,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.24,
      "localCostPer1MUsd": 0.17066,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (350 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (350 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3080-ti",
      "gpuName": "NVIDIA GeForce RTX 3080 Ti",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 93.07,
      "decodeLowTokS": 63.72,
      "decodeHighTokS": 122.41,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 350,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.27,
      "localCostPer1MUsd": 0.15669,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (350 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (350 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3080-ti",
      "gpuName": "NVIDIA GeForce RTX 3080 Ti",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 52.16,
      "decodeLowTokS": 35.71,
      "decodeHighTokS": 68.6,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 350,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.15,
      "localCostPer1MUsd": 0.2796,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (350 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (350 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3080-ti",
      "gpuName": "NVIDIA GeForce RTX 3080 Ti",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 83.1,
      "decodeLowTokS": 56.9,
      "decodeHighTokS": 109.3,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 350,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.24,
      "localCostPer1MUsd": 0.17549,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (350 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (350 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3080-12",
      "gpuName": "NVIDIA GeForce RTX 3080 12GB",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 85.45,
      "decodeLowTokS": 58.51,
      "decodeHighTokS": 112.4,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 350,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.24,
      "localCostPer1MUsd": 0.17066,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (350 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (350 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3080-12",
      "gpuName": "NVIDIA GeForce RTX 3080 12GB",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 93.07,
      "decodeLowTokS": 63.72,
      "decodeHighTokS": 122.41,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 350,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.27,
      "localCostPer1MUsd": 0.15669,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (350 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (350 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3080-12",
      "gpuName": "NVIDIA GeForce RTX 3080 12GB",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 52.16,
      "decodeLowTokS": 35.71,
      "decodeHighTokS": 68.6,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 350,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.15,
      "localCostPer1MUsd": 0.2796,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (350 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (350 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3080-12",
      "gpuName": "NVIDIA GeForce RTX 3080 12GB",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 83.1,
      "decodeLowTokS": 56.9,
      "decodeHighTokS": 109.3,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 350,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.24,
      "localCostPer1MUsd": 0.17549,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (350 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (350 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4060-ti-8",
      "gpuName": "NVIDIA GeForce RTX 4060 Ti 8GB",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 44.75,
      "decodeLowTokS": 30.64,
      "decodeHighTokS": 58.85,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 160,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.28,
      "localCostPer1MUsd": 0.14899,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (160 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (160 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4060-ti-8",
      "gpuName": "NVIDIA GeForce RTX 4060 Ti 8GB",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 49.23,
      "decodeLowTokS": 33.71,
      "decodeHighTokS": 64.75,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 160,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.31,
      "localCostPer1MUsd": 0.13542,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (160 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (160 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4060-ti-8",
      "gpuName": "NVIDIA GeForce RTX 4060 Ti 8GB",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 43.38,
      "decodeLowTokS": 29.7,
      "decodeHighTokS": 57.06,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 160,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.27,
      "localCostPer1MUsd": 0.15368,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (160 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (160 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-7700-xt",
      "gpuName": "AMD Radeon RX 7700 XT",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 44.22,
      "decodeLowTokS": 28.74,
      "decodeHighTokS": 59.7,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 245,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.18,
      "localCostPer1MUsd": 0.23084,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (245 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (245 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-7700-xt",
      "gpuName": "AMD Radeon RX 7700 XT",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 48.36,
      "decodeLowTokS": 31.44,
      "decodeHighTokS": 65.29,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 245,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.2,
      "localCostPer1MUsd": 0.21108,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (245 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (245 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-7700-xt",
      "gpuName": "AMD Radeon RX 7700 XT",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 26.52,
      "decodeLowTokS": 17.24,
      "decodeHighTokS": 35.8,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 245,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.11,
      "localCostPer1MUsd": 0.38497,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (245 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (245 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-7700-xt",
      "gpuName": "AMD Radeon RX 7700 XT",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 42.95,
      "decodeLowTokS": 27.92,
      "decodeHighTokS": 57.99,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 245,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.18,
      "localCostPer1MUsd": 0.23767,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (245 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (245 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4090-laptop",
      "gpuName": "NVIDIA GeForce RTX 4090 Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 81.32,
      "decodeLowTokS": 55.68,
      "decodeHighTokS": 106.96,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 150,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.54,
      "localCostPer1MUsd": 0.07686,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (150 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (150 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4090-laptop",
      "gpuName": "NVIDIA GeForce RTX 4090 Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 88.66,
      "decodeLowTokS": 60.7,
      "decodeHighTokS": 116.61,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 150,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.59,
      "localCostPer1MUsd": 0.0705,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (150 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (150 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4090-laptop",
      "gpuName": "NVIDIA GeForce RTX 4090 Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 49.41,
      "decodeLowTokS": 33.83,
      "decodeHighTokS": 64.99,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 150,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.33,
      "localCostPer1MUsd": 0.12648,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (150 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (150 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4090-laptop",
      "gpuName": "NVIDIA GeForce RTX 4090 Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 79.06,
      "decodeLowTokS": 54.13,
      "decodeHighTokS": 103.98,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 150,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.53,
      "localCostPer1MUsd": 0.07906,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (150 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (150 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4060-laptop",
      "gpuName": "NVIDIA GeForce RTX 4060 Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 40.22,
      "decodeLowTokS": 27.54,
      "decodeHighTokS": 52.91,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 115,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.35,
      "localCostPer1MUsd": 0.11913,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (115 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (115 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4060-laptop",
      "gpuName": "NVIDIA GeForce RTX 4060 Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 44.3,
      "decodeLowTokS": 30.34,
      "decodeHighTokS": 58.27,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 115,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.39,
      "localCostPer1MUsd": 0.10815,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (115 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (115 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4060-laptop",
      "gpuName": "NVIDIA GeForce RTX 4060 Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 38.98,
      "decodeLowTokS": 26.69,
      "decodeHighTokS": 51.27,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 115,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.34,
      "localCostPer1MUsd": 0.12292,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (115 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (115 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "amd-ryzen-ai-9-hx-370",
      "gpuName": "AMD Ryzen AI 9 HX 370",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": true,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 13.83,
      "decodeLowTokS": 8.99,
      "decodeHighTokS": 18.67,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 54,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.26,
      "localCostPer1MUsd": 0.16269,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (54 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (54 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "amd-ryzen-ai-9-hx-370",
      "gpuName": "AMD Ryzen AI 9 HX 370",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": true,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 15.3,
      "decodeLowTokS": 9.95,
      "decodeHighTokS": 20.66,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 54,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.28,
      "localCostPer1MUsd": 0.14701,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (54 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (54 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "amd-ryzen-ai-9-hx-370",
      "gpuName": "AMD Ryzen AI 9 HX 370",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": true,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 7.9,
      "decodeLowTokS": 5.13,
      "decodeHighTokS": 10.66,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 54,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.15,
      "localCostPer1MUsd": 0.28499,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (54 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (54 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "amd-ryzen-ai-9-hx-370",
      "gpuName": "AMD Ryzen AI 9 HX 370",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": true,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 29.24,
      "decodeLowTokS": 19,
      "decodeHighTokS": 39.47,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 54,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.54,
      "localCostPer1MUsd": 0.07696,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (54 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (54 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "amd-ryzen-ai-9-hx-370",
      "gpuName": "AMD Ryzen AI 9 HX 370",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": true,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 13.38,
      "decodeLowTokS": 8.7,
      "decodeHighTokS": 18.07,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 54,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.25,
      "localCostPer1MUsd": 0.16811,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (54 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (54 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-gb10",
      "gpuName": "NVIDIA GB10 Grace Blackwell",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": true,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 39.41,
      "decodeLowTokS": 25.62,
      "decodeHighTokS": 53.2,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 140,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.28,
      "localCostPer1MUsd": 0.14802,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (140 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (140 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-gb10",
      "gpuName": "NVIDIA GB10 Grace Blackwell",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": true,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 43.42,
      "decodeLowTokS": 28.22,
      "decodeHighTokS": 58.61,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 140,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.31,
      "localCostPer1MUsd": 0.13435,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (140 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (140 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-gb10",
      "gpuName": "NVIDIA GB10 Grace Blackwell",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": true,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 22.91,
      "decodeLowTokS": 14.89,
      "decodeHighTokS": 30.93,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 140,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.16,
      "localCostPer1MUsd": 0.25459,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (140 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (140 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-gb10",
      "gpuName": "NVIDIA GB10 Grace Blackwell",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": true,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 79.57,
      "decodeLowTokS": 51.72,
      "decodeHighTokS": 107.42,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 140,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.57,
      "localCostPer1MUsd": 0.07331,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (140 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (140 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-gb10",
      "gpuName": "NVIDIA GB10 Grace Blackwell",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": true,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 38.19,
      "decodeLowTokS": 24.82,
      "decodeHighTokS": 51.56,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 140,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.27,
      "localCostPer1MUsd": 0.15274,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (140 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (140 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m3-max-30c",
      "gpuName": "Apple M3 Max (30-core GPU)",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 28.06,
      "decodeLowTokS": 18.24,
      "decodeHighTokS": 37.88,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.8,
      "localCostPer1MUsd": 0.05197,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m3-max-30c",
      "gpuName": "Apple M3 Max (30-core GPU)",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 30.93,
      "decodeLowTokS": 20.1,
      "decodeHighTokS": 41.75,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.88,
      "localCostPer1MUsd": 0.04715,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m3-max-30c",
      "gpuName": "Apple M3 Max (30-core GPU)",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 16.28,
      "decodeLowTokS": 10.58,
      "decodeHighTokS": 21.98,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.47,
      "localCostPer1MUsd": 0.08958,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m3-max-30c",
      "gpuName": "Apple M3 Max (30-core GPU)",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 56.93,
      "decodeLowTokS": 37,
      "decodeHighTokS": 76.85,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.6,
      "localCostPer1MUsd": 0.02562,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m3-max-30c",
      "gpuName": "Apple M3 Max (30-core GPU)",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 27.19,
      "decodeLowTokS": 17.67,
      "decodeHighTokS": 36.7,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.78,
      "localCostPer1MUsd": 0.05364,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m6",
      "gpuName": "Apple M6",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 16.5,
      "decodeLowTokS": 10.73,
      "decodeHighTokS": 22.28,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 25,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.66,
      "localCostPer1MUsd": 0.06312,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (25 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (25 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m6",
      "gpuName": "Apple M6",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 18.26,
      "decodeLowTokS": 11.87,
      "decodeHighTokS": 24.65,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 25,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.73,
      "localCostPer1MUsd": 0.05705,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (25 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (25 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m6",
      "gpuName": "Apple M6",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 9.43,
      "decodeLowTokS": 6.13,
      "decodeHighTokS": 12.72,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 25,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.38,
      "localCostPer1MUsd": 0.11052,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (25 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (25 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m6",
      "gpuName": "Apple M6",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 34.84,
      "decodeLowTokS": 22.64,
      "decodeHighTokS": 47.03,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 25,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.4,
      "localCostPer1MUsd": 0.0299,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (25 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (25 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m6",
      "gpuName": "Apple M6",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 15.97,
      "decodeLowTokS": 10.38,
      "decodeHighTokS": 21.56,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 25,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.64,
      "localCostPer1MUsd": 0.06522,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (25 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (25 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m5-max-32c",
      "gpuName": "Apple M5 Max (32-core GPU)",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 41.18,
      "decodeLowTokS": 26.76,
      "decodeHighTokS": 55.59,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.2,
      "localCostPer1MUsd": 0.03542,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m5-max-32c",
      "gpuName": "Apple M5 Max (32-core GPU)",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 45.19,
      "decodeLowTokS": 29.37,
      "decodeHighTokS": 61,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.3,
      "localCostPer1MUsd": 0.03227,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m5-max-32c",
      "gpuName": "Apple M5 Max (32-core GPU)",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 24.33,
      "decodeLowTokS": 15.81,
      "decodeHighTokS": 32.84,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.7,
      "localCostPer1MUsd": 0.05994,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m5-max-32c",
      "gpuName": "Apple M5 Max (32-core GPU)",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 80,
      "decodeLowTokS": 52,
      "decodeHighTokS": 108.01,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 2.3,
      "localCostPer1MUsd": 0.01823,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m5-max-32c",
      "gpuName": "Apple M5 Max (32-core GPU)",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 39.95,
      "decodeLowTokS": 25.97,
      "decodeHighTokS": 53.93,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.1,
      "localCostPer1MUsd": 0.0365,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m4-max-32c",
      "gpuName": "Apple M4 Max (32-core GPU)",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 37.2,
      "decodeLowTokS": 24.18,
      "decodeHighTokS": 50.22,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.1,
      "localCostPer1MUsd": 0.0392,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m4-max-32c",
      "gpuName": "Apple M4 Max (32-core GPU)",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 40.88,
      "decodeLowTokS": 26.57,
      "decodeHighTokS": 55.18,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.2,
      "localCostPer1MUsd": 0.03568,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m4-max-32c",
      "gpuName": "Apple M4 Max (32-core GPU)",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 21.86,
      "decodeLowTokS": 14.21,
      "decodeHighTokS": 29.51,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.62,
      "localCostPer1MUsd": 0.06672,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m4-max-32c",
      "gpuName": "Apple M4 Max (32-core GPU)",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 73.22,
      "decodeLowTokS": 47.59,
      "decodeHighTokS": 98.84,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 2.1,
      "localCostPer1MUsd": 0.01992,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m4-max-32c",
      "gpuName": "Apple M4 Max (32-core GPU)",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 36.08,
      "decodeLowTokS": 23.45,
      "decodeHighTokS": 48.7,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1,
      "localCostPer1MUsd": 0.04042,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5090",
      "gpuName": "NVIDIA GeForce RTX 5090",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 173.33,
      "decodeLowTokS": 112.66,
      "decodeHighTokS": 233.99,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 575,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.3,
      "localCostPer1MUsd": 0.13823,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (575 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (575 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5090",
      "gpuName": "NVIDIA GeForce RTX 5090",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 184.75,
      "decodeLowTokS": 120.09,
      "decodeHighTokS": 249.42,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 575,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.32,
      "localCostPer1MUsd": 0.12968,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (575 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (575 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5090",
      "gpuName": "NVIDIA GeForce RTX 5090",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 116.92,
      "decodeLowTokS": 76,
      "decodeHighTokS": 157.84,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 575,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.2,
      "localCostPer1MUsd": 0.20491,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (575 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (575 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5090",
      "gpuName": "NVIDIA GeForce RTX 5090",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 261.89,
      "decodeLowTokS": 170.23,
      "decodeHighTokS": 353.55,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 575,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.46,
      "localCostPer1MUsd": 0.09148,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (575 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (575 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5090",
      "gpuName": "NVIDIA GeForce RTX 5090",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 169.7,
      "decodeLowTokS": 110.31,
      "decodeHighTokS": 229.1,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 575,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.3,
      "localCostPer1MUsd": 0.14118,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (575 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (575 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5080",
      "gpuName": "NVIDIA GeForce RTX 5080",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 113.34,
      "decodeLowTokS": 73.67,
      "decodeHighTokS": 153.01,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 360,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.31,
      "localCostPer1MUsd": 0.13235,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (360 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (360 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5080",
      "gpuName": "NVIDIA GeForce RTX 5080",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 122.59,
      "decodeLowTokS": 79.69,
      "decodeHighTokS": 165.5,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 360,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.34,
      "localCostPer1MUsd": 0.12236,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (360 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (360 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5080",
      "gpuName": "NVIDIA GeForce RTX 5080",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 71.33,
      "decodeLowTokS": 46.37,
      "decodeHighTokS": 96.3,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 360,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.2,
      "localCostPer1MUsd": 0.21028,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (360 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (360 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5080",
      "gpuName": "NVIDIA GeForce RTX 5080",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 110.46,
      "decodeLowTokS": 71.8,
      "decodeHighTokS": 149.12,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 360,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.31,
      "localCostPer1MUsd": 0.1358,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (360 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (360 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5070-ti",
      "gpuName": "NVIDIA GeForce RTX 5070 Ti",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 107.61,
      "decodeLowTokS": 69.95,
      "decodeHighTokS": 145.27,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.36,
      "localCostPer1MUsd": 0.11616,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5070-ti",
      "gpuName": "NVIDIA GeForce RTX 5070 Ti",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 116.56,
      "decodeLowTokS": 75.76,
      "decodeHighTokS": 157.36,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.39,
      "localCostPer1MUsd": 0.10724,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5070-ti",
      "gpuName": "NVIDIA GeForce RTX 5070 Ti",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 67.3,
      "decodeLowTokS": 43.74,
      "decodeHighTokS": 90.85,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.22,
      "localCostPer1MUsd": 0.18575,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5070-ti",
      "gpuName": "NVIDIA GeForce RTX 5070 Ti",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 104.83,
      "decodeLowTokS": 68.14,
      "decodeHighTokS": 141.52,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.35,
      "localCostPer1MUsd": 0.11924,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5070",
      "gpuName": "NVIDIA GeForce RTX 5070",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 85.9,
      "decodeLowTokS": 55.83,
      "decodeHighTokS": 115.96,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 250,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.34,
      "localCostPer1MUsd": 0.12127,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (250 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (250 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5070",
      "gpuName": "NVIDIA GeForce RTX 5070",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 93.54,
      "decodeLowTokS": 60.8,
      "decodeHighTokS": 126.28,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 250,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.37,
      "localCostPer1MUsd": 0.11136,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (250 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (250 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5070",
      "gpuName": "NVIDIA GeForce RTX 5070",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 52.45,
      "decodeLowTokS": 34.1,
      "decodeHighTokS": 70.81,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 250,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.21,
      "localCostPer1MUsd": 0.19859,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (250 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (250 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5070",
      "gpuName": "NVIDIA GeForce RTX 5070",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 83.54,
      "decodeLowTokS": 54.3,
      "decodeHighTokS": 112.78,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 250,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.33,
      "localCostPer1MUsd": 0.12469,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (250 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (250 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5060-ti-16",
      "gpuName": "NVIDIA GeForce RTX 5060 Ti 16GB",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 61.2,
      "decodeLowTokS": 39.78,
      "decodeHighTokS": 82.62,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 180,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.34,
      "localCostPer1MUsd": 0.12255,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (180 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (180 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5060-ti-16",
      "gpuName": "NVIDIA GeForce RTX 5060 Ti 16GB",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 67.06,
      "decodeLowTokS": 43.59,
      "decodeHighTokS": 90.53,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 180,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.37,
      "localCostPer1MUsd": 0.11184,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (180 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (180 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5060-ti-16",
      "gpuName": "NVIDIA GeForce RTX 5060 Ti 16GB",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 36.4,
      "decodeLowTokS": 23.66,
      "decodeHighTokS": 49.14,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 180,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.2,
      "localCostPer1MUsd": 0.20605,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (180 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (180 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5060-ti-16",
      "gpuName": "NVIDIA GeForce RTX 5060 Ti 16GB",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 59.41,
      "decodeLowTokS": 38.61,
      "decodeHighTokS": 80.2,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 180,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.33,
      "localCostPer1MUsd": 0.12625,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (180 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (180 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5060-ti-8",
      "gpuName": "NVIDIA GeForce RTX 5060 Ti 8GB",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 61.2,
      "decodeLowTokS": 39.78,
      "decodeHighTokS": 82.62,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 180,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.34,
      "localCostPer1MUsd": 0.12255,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (180 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (180 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5060-ti-8",
      "gpuName": "NVIDIA GeForce RTX 5060 Ti 8GB",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 67.06,
      "decodeLowTokS": 43.59,
      "decodeHighTokS": 90.53,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 180,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.37,
      "localCostPer1MUsd": 0.11184,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (180 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (180 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5060-ti-8",
      "gpuName": "NVIDIA GeForce RTX 5060 Ti 8GB",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 59.41,
      "decodeLowTokS": 38.61,
      "decodeHighTokS": 80.2,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 180,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.33,
      "localCostPer1MUsd": 0.12625,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (180 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (180 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5060",
      "gpuName": "NVIDIA GeForce RTX 5060",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 61.2,
      "decodeLowTokS": 39.78,
      "decodeHighTokS": 82.62,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 145,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.42,
      "localCostPer1MUsd": 0.09872,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (145 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (145 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5060",
      "gpuName": "NVIDIA GeForce RTX 5060",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 67.06,
      "decodeLowTokS": 43.59,
      "decodeHighTokS": 90.53,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 145,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.46,
      "localCostPer1MUsd": 0.0901,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (145 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (145 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5060",
      "gpuName": "NVIDIA GeForce RTX 5060",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 59.41,
      "decodeLowTokS": 38.61,
      "decodeHighTokS": 80.2,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 145,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.41,
      "localCostPer1MUsd": 0.1017,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (145 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (145 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4070-ti",
      "gpuName": "NVIDIA GeForce RTX 4070 Ti",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 72.82,
      "decodeLowTokS": 49.86,
      "decodeHighTokS": 95.78,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 285,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.26,
      "localCostPer1MUsd": 0.16308,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (285 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (285 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4070-ti",
      "gpuName": "NVIDIA GeForce RTX 4070 Ti",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 79.56,
      "decodeLowTokS": 54.47,
      "decodeHighTokS": 104.64,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 285,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.28,
      "localCostPer1MUsd": 0.14927,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (285 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (285 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4070-ti",
      "gpuName": "NVIDIA GeForce RTX 4070 Ti",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 43.84,
      "decodeLowTokS": 30.02,
      "decodeHighTokS": 57.67,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 285,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.15,
      "localCostPer1MUsd": 0.27084,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (285 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (285 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4070-ti",
      "gpuName": "NVIDIA GeForce RTX 4070 Ti",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 70.75,
      "decodeLowTokS": 48.44,
      "decodeHighTokS": 93.05,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 285,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.25,
      "localCostPer1MUsd": 0.16786,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (285 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (285 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4070-super",
      "gpuName": "NVIDIA GeForce RTX 4070 Super",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 72.82,
      "decodeLowTokS": 49.86,
      "decodeHighTokS": 95.78,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 220,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.33,
      "localCostPer1MUsd": 0.12589,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (220 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (220 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4070-super",
      "gpuName": "NVIDIA GeForce RTX 4070 Super",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 79.56,
      "decodeLowTokS": 54.47,
      "decodeHighTokS": 104.64,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 220,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.36,
      "localCostPer1MUsd": 0.11522,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (220 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (220 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4070-super",
      "gpuName": "NVIDIA GeForce RTX 4070 Super",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 43.84,
      "decodeLowTokS": 30.02,
      "decodeHighTokS": 57.67,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 220,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.2,
      "localCostPer1MUsd": 0.20907,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (220 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (220 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4070-super",
      "gpuName": "NVIDIA GeForce RTX 4070 Super",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 70.75,
      "decodeLowTokS": 48.44,
      "decodeHighTokS": 93.05,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 220,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.32,
      "localCostPer1MUsd": 0.12957,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (220 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (220 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4090",
      "gpuName": "NVIDIA GeForce RTX 4090",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 125.16,
      "decodeLowTokS": 85.7,
      "decodeHighTokS": 164.63,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 450,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.28,
      "localCostPer1MUsd": 0.1498,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (450 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (450 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4090",
      "gpuName": "NVIDIA GeForce RTX 4090",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 135,
      "decodeLowTokS": 135,
      "decodeHighTokS": 135,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 450,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.3,
      "localCostPer1MUsd": 0.13889,
      "provenance": {
        "decodeBadge": "MEASURED · 1 run",
        "decodeTier": "measured",
        "tokensPerWattBadge": "ESTIMATED · spec board power (TGP/TBP)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · spec board power (TGP/TBP)",
        "localCostTier": "estimated",
        "mixed": true,
        "offloadWarning": false,
        "speed": {
          "tier": "measured",
          "runs": 1,
          "asOf": "2026-08-17"
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (450 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (450 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4090",
      "gpuName": "NVIDIA GeForce RTX 4090",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 79.83,
      "decodeLowTokS": 54.66,
      "decodeHighTokS": 105,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 450,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.18,
      "localCostPer1MUsd": 0.23488,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (450 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (450 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4090",
      "gpuName": "NVIDIA GeForce RTX 4090",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 207.94,
      "decodeLowTokS": 142.38,
      "decodeHighTokS": 273.5,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 450,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.46,
      "localCostPer1MUsd": 0.09017,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (450 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (450 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4090",
      "gpuName": "NVIDIA GeForce RTX 4090",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 141.3,
      "decodeLowTokS": 141.3,
      "decodeHighTokS": 141.3,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 450,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.31,
      "localCostPer1MUsd": 0.1327,
      "provenance": {
        "decodeBadge": "MEASURED · 1 run",
        "decodeTier": "measured",
        "tokensPerWattBadge": "ESTIMATED · spec board power (TGP/TBP)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · spec board power (TGP/TBP)",
        "localCostTier": "estimated",
        "mixed": true,
        "offloadWarning": false,
        "speed": {
          "tier": "measured",
          "runs": 1,
          "asOf": "2026-08-17"
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (450 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (450 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4080-super",
      "gpuName": "NVIDIA GeForce RTX 4080 Super",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 98.89,
      "decodeLowTokS": 67.71,
      "decodeHighTokS": 130.07,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 320,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.31,
      "localCostPer1MUsd": 0.13483,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (320 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (320 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4080-super",
      "gpuName": "NVIDIA GeForce RTX 4080 Super",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 107.35,
      "decodeLowTokS": 73.5,
      "decodeHighTokS": 141.19,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 320,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.34,
      "localCostPer1MUsd": 0.12421,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (320 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (320 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4080-super",
      "gpuName": "NVIDIA GeForce RTX 4080 Super",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 61.25,
      "decodeLowTokS": 41.94,
      "decodeHighTokS": 80.56,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 320,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.19,
      "localCostPer1MUsd": 0.21768,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (320 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (320 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4080-super",
      "gpuName": "NVIDIA GeForce RTX 4080 Super",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 96.27,
      "decodeLowTokS": 65.92,
      "decodeHighTokS": 126.62,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 320,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.3,
      "localCostPer1MUsd": 0.1385,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (320 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (320 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4080",
      "gpuName": "NVIDIA GeForce RTX 4080",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 96.89,
      "decodeLowTokS": 66.34,
      "decodeHighTokS": 127.44,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 320,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.3,
      "localCostPer1MUsd": 0.13761,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (320 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (320 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4080",
      "gpuName": "NVIDIA GeForce RTX 4080",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 105.23,
      "decodeLowTokS": 72.05,
      "decodeHighTokS": 138.41,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 320,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.33,
      "localCostPer1MUsd": 0.12671,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (320 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (320 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4080",
      "gpuName": "NVIDIA GeForce RTX 4080",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 59.88,
      "decodeLowTokS": 41,
      "decodeHighTokS": 78.76,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 320,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.19,
      "localCostPer1MUsd": 0.22266,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (320 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (320 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4080",
      "gpuName": "NVIDIA GeForce RTX 4080",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 94.31,
      "decodeLowTokS": 64.57,
      "decodeHighTokS": 124.05,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 320,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.29,
      "localCostPer1MUsd": 0.14138,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (320 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (320 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4070-ti-super",
      "gpuName": "NVIDIA GeForce RTX 4070 Ti Super",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 92.07,
      "decodeLowTokS": 63.04,
      "decodeHighTokS": 121.1,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 285,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.32,
      "localCostPer1MUsd": 0.12898,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (285 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (285 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4070-ti-super",
      "gpuName": "NVIDIA GeForce RTX 4070 Ti Super",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 100.11,
      "decodeLowTokS": 68.55,
      "decodeHighTokS": 131.68,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 285,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.35,
      "localCostPer1MUsd": 0.11862,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (285 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (285 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4070-ti-super",
      "gpuName": "NVIDIA GeForce RTX 4070 Ti Super",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 56.6,
      "decodeLowTokS": 38.76,
      "decodeHighTokS": 74.45,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 285,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.2,
      "localCostPer1MUsd": 0.2098,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (285 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (285 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4070-ti-super",
      "gpuName": "NVIDIA GeForce RTX 4070 Ti Super",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 89.58,
      "decodeLowTokS": 61.34,
      "decodeHighTokS": 117.83,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 285,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.31,
      "localCostPer1MUsd": 0.13256,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (285 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (285 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4070",
      "gpuName": "NVIDIA GeForce RTX 4070",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 72.82,
      "decodeLowTokS": 49.86,
      "decodeHighTokS": 95.78,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 200,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.36,
      "localCostPer1MUsd": 0.11444,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (200 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (200 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4070",
      "gpuName": "NVIDIA GeForce RTX 4070",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 79.56,
      "decodeLowTokS": 54.47,
      "decodeHighTokS": 104.64,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 200,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.4,
      "localCostPer1MUsd": 0.10475,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (200 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (200 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4070",
      "gpuName": "NVIDIA GeForce RTX 4070",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 43.84,
      "decodeLowTokS": 30.02,
      "decodeHighTokS": 57.67,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 200,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.22,
      "localCostPer1MUsd": 0.19006,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (200 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (200 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4070",
      "gpuName": "NVIDIA GeForce RTX 4070",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 70.75,
      "decodeLowTokS": 48.44,
      "decodeHighTokS": 93.05,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 200,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.35,
      "localCostPer1MUsd": 0.11779,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (200 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (200 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4060-ti-16",
      "gpuName": "NVIDIA GeForce RTX 4060 Ti 16GB",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 44.75,
      "decodeLowTokS": 30.64,
      "decodeHighTokS": 58.85,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 165,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.27,
      "localCostPer1MUsd": 0.15365,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (165 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (165 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4060-ti-16",
      "gpuName": "NVIDIA GeForce RTX 4060 Ti 16GB",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 49.23,
      "decodeLowTokS": 33.71,
      "decodeHighTokS": 64.75,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 165,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.3,
      "localCostPer1MUsd": 0.13965,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (165 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (165 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4060-ti-16",
      "gpuName": "NVIDIA GeForce RTX 4060 Ti 16GB",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 26.16,
      "decodeLowTokS": 17.91,
      "decodeHighTokS": 34.41,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 165,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.16,
      "localCostPer1MUsd": 0.26282,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (165 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (165 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4060-ti-16",
      "gpuName": "NVIDIA GeForce RTX 4060 Ti 16GB",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 43.38,
      "decodeLowTokS": 29.7,
      "decodeHighTokS": 57.06,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 165,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.26,
      "localCostPer1MUsd": 0.15848,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (165 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (165 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4060",
      "gpuName": "NVIDIA GeForce RTX 4060",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 42.5,
      "decodeLowTokS": 29.1,
      "decodeHighTokS": 55.9,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 115,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.37,
      "localCostPer1MUsd": 0.11275,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (115 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (115 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4060",
      "gpuName": "NVIDIA GeForce RTX 4060",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 46.78,
      "decodeLowTokS": 32.03,
      "decodeHighTokS": 61.53,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 115,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.41,
      "localCostPer1MUsd": 0.10242,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (115 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (115 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-4060",
      "gpuName": "NVIDIA GeForce RTX 4060",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 41.19,
      "decodeLowTokS": 28.2,
      "decodeHighTokS": 54.18,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 115,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.36,
      "localCostPer1MUsd": 0.11632,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (115 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (115 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3070",
      "gpuName": "NVIDIA GeForce RTX 3070",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 46.52,
      "decodeLowTokS": 31.85,
      "decodeHighTokS": 61.19,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 220,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.21,
      "localCostPer1MUsd": 0.19705,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (220 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (220 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3070",
      "gpuName": "NVIDIA GeForce RTX 3070",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 51.16,
      "decodeLowTokS": 35.03,
      "decodeHighTokS": 67.29,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 220,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.23,
      "localCostPer1MUsd": 0.17918,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (220 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (220 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3070",
      "gpuName": "NVIDIA GeForce RTX 3070",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 45.11,
      "decodeLowTokS": 30.88,
      "decodeHighTokS": 59.33,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 220,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.21,
      "localCostPer1MUsd": 0.20323,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (220 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (220 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3070-ti",
      "gpuName": "NVIDIA GeForce RTX 3070 Ti",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 60.86,
      "decodeLowTokS": 41.67,
      "decodeHighTokS": 80.05,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 290,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.21,
      "localCostPer1MUsd": 0.19854,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (290 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (290 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3070-ti",
      "gpuName": "NVIDIA GeForce RTX 3070 Ti",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 66.69,
      "decodeLowTokS": 45.66,
      "decodeHighTokS": 87.72,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 290,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.23,
      "localCostPer1MUsd": 0.18118,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (290 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (290 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3070-ti",
      "gpuName": "NVIDIA GeForce RTX 3070 Ti",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 59.08,
      "decodeLowTokS": 40.45,
      "decodeHighTokS": 77.7,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 290,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.2,
      "localCostPer1MUsd": 0.20454,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (290 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (290 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3090",
      "gpuName": "NVIDIA GeForce RTX 3090",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 87.26,
      "decodeLowTokS": 59.75,
      "decodeHighTokS": 114.77,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 350,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.25,
      "localCostPer1MUsd": 0.16712,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (350 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (350 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3090",
      "gpuName": "NVIDIA GeForce RTX 3090",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 95,
      "decodeLowTokS": 95,
      "decodeHighTokS": 95,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 350,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.27,
      "localCostPer1MUsd": 0.15351,
      "provenance": {
        "decodeBadge": "MEASURED · 1 run",
        "decodeTier": "measured",
        "tokensPerWattBadge": "ESTIMATED · spec board power (TGP/TBP)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · spec board power (TGP/TBP)",
        "localCostTier": "estimated",
        "mixed": true,
        "offloadWarning": false,
        "speed": {
          "tier": "measured",
          "runs": 1,
          "asOf": "2026-08-17"
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (350 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (350 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3090",
      "gpuName": "NVIDIA GeForce RTX 3090",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 53.37,
      "decodeLowTokS": 36.54,
      "decodeHighTokS": 70.19,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 350,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.15,
      "localCostPer1MUsd": 0.27327,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (350 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (350 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3090",
      "gpuName": "NVIDIA GeForce RTX 3090",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 157.29,
      "decodeLowTokS": 107.69,
      "decodeHighTokS": 206.88,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 350,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.45,
      "localCostPer1MUsd": 0.09272,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (350 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (350 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3090",
      "gpuName": "NVIDIA GeForce RTX 3090",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 84.87,
      "decodeLowTokS": 58.11,
      "decodeHighTokS": 111.63,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 350,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.24,
      "localCostPer1MUsd": 0.17183,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (350 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (350 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3080",
      "gpuName": "NVIDIA GeForce RTX 3080 (10GB)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 73.56,
      "decodeLowTokS": 50.37,
      "decodeHighTokS": 96.76,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 320,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.23,
      "localCostPer1MUsd": 0.18125,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (320 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (320 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3080",
      "gpuName": "NVIDIA GeForce RTX 3080 (10GB)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 80.36,
      "decodeLowTokS": 55.02,
      "decodeHighTokS": 105.69,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 320,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.25,
      "localCostPer1MUsd": 0.16593,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (320 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (320 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3080",
      "gpuName": "NVIDIA GeForce RTX 3080 (10GB)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 71.47,
      "decodeLowTokS": 48.94,
      "decodeHighTokS": 94.01,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 320,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.22,
      "localCostPer1MUsd": 0.18655,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (320 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (320 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3060",
      "gpuName": "NVIDIA GeForce RTX 3060 (12GB)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 38.16,
      "decodeLowTokS": 26.13,
      "decodeHighTokS": 50.2,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 170,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.22,
      "localCostPer1MUsd": 0.1856,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (170 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (170 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3060",
      "gpuName": "NVIDIA GeForce RTX 3060 (12GB)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 42.06,
      "decodeLowTokS": 28.8,
      "decodeHighTokS": 55.32,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 170,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.25,
      "localCostPer1MUsd": 0.16842,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (170 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (170 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3060",
      "gpuName": "NVIDIA GeForce RTX 3060 (12GB)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 22.16,
      "decodeLowTokS": 15.17,
      "decodeHighTokS": 29.15,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 170,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.13,
      "localCostPer1MUsd": 0.31965,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (170 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (170 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3060",
      "gpuName": "NVIDIA GeForce RTX 3060 (12GB)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 36.98,
      "decodeLowTokS": 25.32,
      "decodeHighTokS": 48.64,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 170,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.22,
      "localCostPer1MUsd": 0.19154,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (170 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (170 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-7900-xtx",
      "gpuName": "AMD Radeon RX 7900 XTX",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 82.64,
      "decodeLowTokS": 53.72,
      "decodeHighTokS": 111.56,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 355,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.23,
      "localCostPer1MUsd": 0.17899,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (355 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (355 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-7900-xtx",
      "gpuName": "AMD Radeon RX 7900 XTX",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 89.05,
      "decodeLowTokS": 57.88,
      "decodeHighTokS": 120.22,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 355,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.25,
      "localCostPer1MUsd": 0.16611,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (355 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (355 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-7900-xtx",
      "gpuName": "AMD Radeon RX 7900 XTX",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 52.92,
      "decodeLowTokS": 34.4,
      "decodeHighTokS": 71.45,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 355,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.15,
      "localCostPer1MUsd": 0.27949,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (355 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (355 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-7900-xtx",
      "gpuName": "AMD Radeon RX 7900 XTX",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 136.27,
      "decodeLowTokS": 88.58,
      "decodeHighTokS": 183.97,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 355,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.38,
      "localCostPer1MUsd": 0.10854,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (355 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (355 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-7900-xtx",
      "gpuName": "AMD Radeon RX 7900 XTX",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 80.63,
      "decodeLowTokS": 52.41,
      "decodeHighTokS": 108.86,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 355,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.23,
      "localCostPer1MUsd": 0.18344,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (355 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (355 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-7900-xt",
      "gpuName": "AMD Radeon RX 7900 XT",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 72.35,
      "decodeLowTokS": 47.03,
      "decodeHighTokS": 97.68,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 315,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.23,
      "localCostPer1MUsd": 0.1814,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (315 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (315 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-7900-xt",
      "gpuName": "AMD Radeon RX 7900 XT",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 78.27,
      "decodeLowTokS": 50.88,
      "decodeHighTokS": 105.67,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 315,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.25,
      "localCostPer1MUsd": 0.16768,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (315 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (315 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-7900-xt",
      "gpuName": "AMD Radeon RX 7900 XT",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 45.51,
      "decodeLowTokS": 29.58,
      "decodeHighTokS": 61.44,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 315,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.14,
      "localCostPer1MUsd": 0.28841,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (315 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (315 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-7900-xt",
      "gpuName": "AMD Radeon RX 7900 XT",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 123.37,
      "decodeLowTokS": 80.19,
      "decodeHighTokS": 166.55,
      "requiredVramGb": 19.62,
      "fitStatus": "tight",
      "watts": 315,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.39,
      "localCostPer1MUsd": 0.10639,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (315 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (315 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-7900-xt",
      "gpuName": "AMD Radeon RX 7900 XT",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 70.51,
      "decodeLowTokS": 45.83,
      "decodeHighTokS": 95.19,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 315,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.22,
      "localCostPer1MUsd": 0.18614,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (315 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (315 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-7800-xt",
      "gpuName": "AMD Radeon RX 7800 XT",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 59.77,
      "decodeLowTokS": 38.85,
      "decodeHighTokS": 80.68,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 263,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.23,
      "localCostPer1MUsd": 0.18335,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (263 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (263 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-7800-xt",
      "gpuName": "AMD Radeon RX 7800 XT",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 64.97,
      "decodeLowTokS": 42.23,
      "decodeHighTokS": 87.71,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 263,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.25,
      "localCostPer1MUsd": 0.16867,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (263 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (263 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-7800-xt",
      "gpuName": "AMD Radeon RX 7800 XT",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 36.79,
      "decodeLowTokS": 23.91,
      "decodeHighTokS": 49.66,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 263,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.14,
      "localCostPer1MUsd": 0.2979,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (263 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (263 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-7800-xt",
      "gpuName": "AMD Radeon RX 7800 XT",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 58.16,
      "decodeLowTokS": 37.8,
      "decodeHighTokS": 78.51,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 263,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.22,
      "localCostPer1MUsd": 0.18843,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (263 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (263 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-9070-xt",
      "gpuName": "AMD Radeon RX 9070 XT",
      "vendor": "AMD",
      "architecture": "amd-rdna4",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 64.6,
      "decodeLowTokS": 41.99,
      "decodeHighTokS": 87.21,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 304,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.21,
      "localCostPer1MUsd": 0.19608,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna4",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (304 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (304 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-9070-xt",
      "gpuName": "AMD Radeon RX 9070 XT",
      "vendor": "AMD",
      "architecture": "amd-rdna4",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 70.09,
      "decodeLowTokS": 45.56,
      "decodeHighTokS": 94.63,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 304,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.23,
      "localCostPer1MUsd": 0.18071,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna4",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (304 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (304 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-9070-xt",
      "gpuName": "AMD Radeon RX 9070 XT",
      "vendor": "AMD",
      "architecture": "amd-rdna4",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 40.09,
      "decodeLowTokS": 26.06,
      "decodeHighTokS": 54.12,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 304,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.13,
      "localCostPer1MUsd": 0.31595,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna4",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (304 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (304 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-9070-xt",
      "gpuName": "AMD Radeon RX 9070 XT",
      "vendor": "AMD",
      "architecture": "amd-rdna4",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 62.9,
      "decodeLowTokS": 40.88,
      "decodeHighTokS": 84.91,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 304,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.21,
      "localCostPer1MUsd": 0.20139,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna4",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (304 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (304 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-9070",
      "gpuName": "AMD Radeon RX 9070",
      "vendor": "AMD",
      "architecture": "amd-rdna4",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 64.6,
      "decodeLowTokS": 41.99,
      "decodeHighTokS": 87.21,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 220,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.29,
      "localCostPer1MUsd": 0.1419,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna4",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (220 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (220 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-9070",
      "gpuName": "AMD Radeon RX 9070",
      "vendor": "AMD",
      "architecture": "amd-rdna4",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 70.09,
      "decodeLowTokS": 45.56,
      "decodeHighTokS": 94.63,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 220,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.32,
      "localCostPer1MUsd": 0.13078,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna4",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (220 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (220 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-9070",
      "gpuName": "AMD Radeon RX 9070",
      "vendor": "AMD",
      "architecture": "amd-rdna4",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 40.09,
      "decodeLowTokS": 26.06,
      "decodeHighTokS": 54.12,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 220,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.18,
      "localCostPer1MUsd": 0.22865,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna4",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (220 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (220 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-9070",
      "gpuName": "AMD Radeon RX 9070",
      "vendor": "AMD",
      "architecture": "amd-rdna4",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 62.9,
      "decodeLowTokS": 40.88,
      "decodeHighTokS": 84.91,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 220,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.29,
      "localCostPer1MUsd": 0.14574,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna4",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (220 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (220 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-9060-xt-16",
      "gpuName": "AMD Radeon RX 9060 XT 16GB",
      "vendor": "AMD",
      "architecture": "amd-rdna4",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 36.42,
      "decodeLowTokS": 23.67,
      "decodeHighTokS": 49.16,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 160,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.23,
      "localCostPer1MUsd": 0.18306,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna4",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (160 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (160 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-9060-xt-16",
      "gpuName": "AMD Radeon RX 9060 XT 16GB",
      "vendor": "AMD",
      "architecture": "amd-rdna4",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 39.95,
      "decodeLowTokS": 25.97,
      "decodeHighTokS": 53.93,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 160,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.25,
      "localCostPer1MUsd": 0.16689,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna4",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (160 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (160 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-9060-xt-16",
      "gpuName": "AMD Radeon RX 9060 XT 16GB",
      "vendor": "AMD",
      "architecture": "amd-rdna4",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 21.56,
      "decodeLowTokS": 14.01,
      "decodeHighTokS": 29.1,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 160,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.13,
      "localCostPer1MUsd": 0.30924,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna4",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (160 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (160 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-9060-xt-16",
      "gpuName": "AMD Radeon RX 9060 XT 16GB",
      "vendor": "AMD",
      "architecture": "amd-rdna4",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 35.34,
      "decodeLowTokS": 22.97,
      "decodeHighTokS": 47.71,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 160,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.22,
      "localCostPer1MUsd": 0.18865,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna4",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (160 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (160 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-9060-xt-8",
      "gpuName": "AMD Radeon RX 9060 XT 8GB",
      "vendor": "AMD",
      "architecture": "amd-rdna4",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 36.42,
      "decodeLowTokS": 23.67,
      "decodeHighTokS": 49.16,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 150,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.24,
      "localCostPer1MUsd": 0.17162,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna4",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (150 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (150 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-9060-xt-8",
      "gpuName": "AMD Radeon RX 9060 XT 8GB",
      "vendor": "AMD",
      "architecture": "amd-rdna4",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 39.95,
      "decodeLowTokS": 25.97,
      "decodeHighTokS": 53.93,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 150,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.27,
      "localCostPer1MUsd": 0.15646,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna4",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (150 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (150 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-9060-xt-8",
      "gpuName": "AMD Radeon RX 9060 XT 8GB",
      "vendor": "AMD",
      "architecture": "amd-rdna4",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 35.34,
      "decodeLowTokS": 22.97,
      "decodeHighTokS": 47.71,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 150,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.24,
      "localCostPer1MUsd": 0.17686,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna4",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (150 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (150 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m1-max",
      "gpuName": "Apple M1 Max",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 36.39,
      "decodeLowTokS": 23.65,
      "decodeHighTokS": 49.13,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 30,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.2,
      "localCostPer1MUsd": 0.03435,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (30 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (30 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m1-max",
      "gpuName": "Apple M1 Max",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 40,
      "decodeLowTokS": 26,
      "decodeHighTokS": 54,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 30,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.3,
      "localCostPer1MUsd": 0.03125,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (30 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (30 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m1-max",
      "gpuName": "Apple M1 Max",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 21.36,
      "decodeLowTokS": 13.88,
      "decodeHighTokS": 28.84,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 30,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.71,
      "localCostPer1MUsd": 0.05852,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (30 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (30 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m1-max",
      "gpuName": "Apple M1 Max",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 71.82,
      "decodeLowTokS": 46.68,
      "decodeHighTokS": 96.95,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 30,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 2.4,
      "localCostPer1MUsd": 0.01741,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (30 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (30 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m1-max",
      "gpuName": "Apple M1 Max",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 35.29,
      "decodeLowTokS": 22.94,
      "decodeHighTokS": 47.64,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 30,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.2,
      "localCostPer1MUsd": 0.03542,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (30 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (30 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m2-pro",
      "gpuName": "Apple M2 Pro",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 19.25,
      "decodeLowTokS": 12.51,
      "decodeHighTokS": 25.98,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 30,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.64,
      "localCostPer1MUsd": 0.06495,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (30 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (30 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m2-pro",
      "gpuName": "Apple M2 Pro",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 21.28,
      "decodeLowTokS": 13.83,
      "decodeHighTokS": 28.72,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 30,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.71,
      "localCostPer1MUsd": 0.05875,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (30 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (30 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m2-pro",
      "gpuName": "Apple M2 Pro",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 11.03,
      "decodeLowTokS": 7.17,
      "decodeHighTokS": 14.89,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 30,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.37,
      "localCostPer1MUsd": 0.11329,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (30 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (30 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m2-pro",
      "gpuName": "Apple M2 Pro",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 40.24,
      "decodeLowTokS": 26.16,
      "decodeHighTokS": 54.33,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 30,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.3,
      "localCostPer1MUsd": 0.03106,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (30 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (30 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m2-pro",
      "gpuName": "Apple M2 Pro",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 18.63,
      "decodeLowTokS": 12.11,
      "decodeHighTokS": 25.15,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 30,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.62,
      "localCostPer1MUsd": 0.06709,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (30 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (30 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m2-max",
      "gpuName": "Apple M2 Max",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 36.39,
      "decodeLowTokS": 23.65,
      "decodeHighTokS": 49.13,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 40,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.91,
      "localCostPer1MUsd": 0.0458,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (40 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (40 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m2-max",
      "gpuName": "Apple M2 Max",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 40,
      "decodeLowTokS": 26,
      "decodeHighTokS": 54,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 40,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1,
      "localCostPer1MUsd": 0.04167,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (40 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (40 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m2-max",
      "gpuName": "Apple M2 Max",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 21.36,
      "decodeLowTokS": 13.88,
      "decodeHighTokS": 28.84,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 40,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.53,
      "localCostPer1MUsd": 0.07803,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (40 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (40 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m2-max",
      "gpuName": "Apple M2 Max",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 71.82,
      "decodeLowTokS": 46.68,
      "decodeHighTokS": 96.95,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 40,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.8,
      "localCostPer1MUsd": 0.02321,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (40 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (40 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m2-max",
      "gpuName": "Apple M2 Max",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 35.29,
      "decodeLowTokS": 22.94,
      "decodeHighTokS": 47.64,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 40,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.88,
      "localCostPer1MUsd": 0.04723,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (40 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (40 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m3-pro",
      "gpuName": "Apple M3 Pro",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 14.93,
      "decodeLowTokS": 9.7,
      "decodeHighTokS": 20.15,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 30,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.5,
      "localCostPer1MUsd": 0.08375,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (30 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (30 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m3-pro",
      "gpuName": "Apple M3 Pro",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 16.52,
      "decodeLowTokS": 10.74,
      "decodeHighTokS": 22.31,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 30,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.55,
      "localCostPer1MUsd": 0.07565,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (30 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (30 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m3-pro",
      "gpuName": "Apple M3 Pro",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 8.51,
      "decodeLowTokS": 5.53,
      "decodeHighTokS": 11.48,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 30,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.28,
      "localCostPer1MUsd": 0.14695,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (30 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (30 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m3-pro",
      "gpuName": "Apple M3 Pro",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 31.69,
      "decodeLowTokS": 20.6,
      "decodeHighTokS": 42.78,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 30,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.1,
      "localCostPer1MUsd": 0.03945,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (30 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (30 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m3-pro",
      "gpuName": "Apple M3 Pro",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 14.44,
      "decodeLowTokS": 9.39,
      "decodeHighTokS": 19.5,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 30,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.48,
      "localCostPer1MUsd": 0.08655,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (30 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (30 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m4-max",
      "gpuName": "Apple M4 Max",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 47.77,
      "decodeLowTokS": 31.05,
      "decodeHighTokS": 64.49,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.4,
      "localCostPer1MUsd": 0.03053,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m4-max",
      "gpuName": "Apple M4 Max",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 52.31,
      "decodeLowTokS": 34,
      "decodeHighTokS": 70.62,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.5,
      "localCostPer1MUsd": 0.02788,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m4-max",
      "gpuName": "Apple M4 Max",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 28.49,
      "decodeLowTokS": 18.52,
      "decodeHighTokS": 38.46,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.81,
      "localCostPer1MUsd": 0.05119,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m4-max",
      "gpuName": "Apple M4 Max",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 90.88,
      "decodeLowTokS": 59.07,
      "decodeHighTokS": 122.69,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 2.6,
      "localCostPer1MUsd": 0.01605,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m4-max",
      "gpuName": "Apple M4 Max",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 46.38,
      "decodeLowTokS": 30.15,
      "decodeHighTokS": 62.61,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.3,
      "localCostPer1MUsd": 0.03144,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m4-pro",
      "gpuName": "Apple M4 Pro",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 25.73,
      "decodeLowTokS": 16.72,
      "decodeHighTokS": 34.73,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 30,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.86,
      "localCostPer1MUsd": 0.04858,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (30 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (30 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m4-pro",
      "gpuName": "Apple M4 Pro",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 28.38,
      "decodeLowTokS": 18.45,
      "decodeHighTokS": 38.31,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 30,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.95,
      "localCostPer1MUsd": 0.04404,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (30 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (30 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m4-pro",
      "gpuName": "Apple M4 Pro",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 14.88,
      "decodeLowTokS": 9.67,
      "decodeHighTokS": 20.09,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 30,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.5,
      "localCostPer1MUsd": 0.084,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (30 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (30 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m4-pro",
      "gpuName": "Apple M4 Pro",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 24.92,
      "decodeLowTokS": 16.2,
      "decodeHighTokS": 33.65,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 30,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.83,
      "localCostPer1MUsd": 0.05015,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (30 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (30 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m3-max",
      "gpuName": "Apple M3 Max",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 36.39,
      "decodeLowTokS": 23.65,
      "decodeHighTokS": 49.13,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1,
      "localCostPer1MUsd": 0.04007,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m3-max",
      "gpuName": "Apple M3 Max",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 40,
      "decodeLowTokS": 40,
      "decodeHighTokS": 40,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.1,
      "localCostPer1MUsd": 0.03646,
      "provenance": {
        "decodeBadge": "MEASURED · 1 run",
        "decodeTier": "measured",
        "tokensPerWattBadge": "ESTIMATED · spec SoC package power",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · spec SoC package power",
        "localCostTier": "estimated",
        "mixed": true,
        "offloadWarning": false,
        "speed": {
          "tier": "measured",
          "runs": 1,
          "asOf": "2026-08-17"
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m3-max",
      "gpuName": "Apple M3 Max",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 21.36,
      "decodeLowTokS": 13.88,
      "decodeHighTokS": 28.84,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.61,
      "localCostPer1MUsd": 0.06828,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m3-max",
      "gpuName": "Apple M3 Max",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 71.82,
      "decodeLowTokS": 46.68,
      "decodeHighTokS": 96.95,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 2.1,
      "localCostPer1MUsd": 0.02031,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m3-max",
      "gpuName": "Apple M3 Max",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 35.29,
      "decodeLowTokS": 22.94,
      "decodeHighTokS": 47.64,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1,
      "localCostPer1MUsd": 0.04132,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m1",
      "gpuName": "Apple M1",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 6.8,
      "decodeLowTokS": 4.42,
      "decodeHighTokS": 9.18,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 20,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.34,
      "localCostPer1MUsd": 0.1225,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (20 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (20 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m1",
      "gpuName": "Apple M1",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 7.55,
      "decodeLowTokS": 4.91,
      "decodeHighTokS": 10.2,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 20,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.38,
      "localCostPer1MUsd": 0.11035,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (20 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (20 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m1",
      "gpuName": "Apple M1",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 3.84,
      "decodeLowTokS": 2.49,
      "decodeHighTokS": 5.18,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 20,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.19,
      "localCostPer1MUsd": 0.21729,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (20 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (20 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m1",
      "gpuName": "Apple M1",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 6.58,
      "decodeLowTokS": 4.28,
      "decodeHighTokS": 8.88,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 20,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.33,
      "localCostPer1MUsd": 0.1267,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (20 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (20 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m1-pro",
      "gpuName": "Apple M1 Pro",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 19.25,
      "decodeLowTokS": 12.51,
      "decodeHighTokS": 25.98,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 30,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.64,
      "localCostPer1MUsd": 0.06495,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (30 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (30 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m1-pro",
      "gpuName": "Apple M1 Pro",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 21.28,
      "decodeLowTokS": 13.83,
      "decodeHighTokS": 28.72,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 30,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.71,
      "localCostPer1MUsd": 0.05875,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (30 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (30 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m1-pro",
      "gpuName": "Apple M1 Pro",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 11.03,
      "decodeLowTokS": 7.17,
      "decodeHighTokS": 14.89,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 30,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.37,
      "localCostPer1MUsd": 0.11329,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (30 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (30 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m1-pro",
      "gpuName": "Apple M1 Pro",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 40.24,
      "decodeLowTokS": 26.16,
      "decodeHighTokS": 54.33,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 30,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.3,
      "localCostPer1MUsd": 0.03106,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (30 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (30 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m1-pro",
      "gpuName": "Apple M1 Pro",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 18.63,
      "decodeLowTokS": 12.11,
      "decodeHighTokS": 25.15,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 30,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.62,
      "localCostPer1MUsd": 0.06709,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (30 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (30 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m1-ultra",
      "gpuName": "Apple M1 Ultra",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 65.62,
      "decodeLowTokS": 42.65,
      "decodeHighTokS": 88.58,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 60,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.1,
      "localCostPer1MUsd": 0.0381,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (60 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (60 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m1-ultra",
      "gpuName": "Apple M1 Ultra",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 71.43,
      "decodeLowTokS": 46.43,
      "decodeHighTokS": 96.43,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 60,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.2,
      "localCostPer1MUsd": 0.035,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (60 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (60 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m1-ultra",
      "gpuName": "Apple M1 Ultra",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 40.15,
      "decodeLowTokS": 26.1,
      "decodeHighTokS": 54.2,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 60,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.67,
      "localCostPer1MUsd": 0.06227,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (60 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (60 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m1-ultra",
      "gpuName": "Apple M1 Ultra",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 118.17,
      "decodeLowTokS": 76.81,
      "decodeHighTokS": 159.53,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 60,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 2,
      "localCostPer1MUsd": 0.02116,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (60 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (60 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m1-ultra",
      "gpuName": "Apple M1 Ultra",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 63.82,
      "decodeLowTokS": 41.49,
      "decodeHighTokS": 86.16,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 60,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.1,
      "localCostPer1MUsd": 0.03917,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (60 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (60 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m2",
      "gpuName": "Apple M2",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 9.91,
      "decodeLowTokS": 6.44,
      "decodeHighTokS": 13.38,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 20,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.5,
      "localCostPer1MUsd": 0.0841,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (20 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (20 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m2",
      "gpuName": "Apple M2",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 10.99,
      "decodeLowTokS": 7.14,
      "decodeHighTokS": 14.83,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 20,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.55,
      "localCostPer1MUsd": 0.07583,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (20 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (20 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m2",
      "gpuName": "Apple M2",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 5.61,
      "decodeLowTokS": 3.65,
      "decodeHighTokS": 7.57,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 20,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.28,
      "localCostPer1MUsd": 0.14856,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (20 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (20 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m2",
      "gpuName": "Apple M2",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 9.58,
      "decodeLowTokS": 6.23,
      "decodeHighTokS": 12.94,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 20,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.48,
      "localCostPer1MUsd": 0.08695,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (20 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (20 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m2-ultra",
      "gpuName": "Apple M2 Ultra",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 65.62,
      "decodeLowTokS": 42.65,
      "decodeHighTokS": 88.58,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 60,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.1,
      "localCostPer1MUsd": 0.0381,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (60 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (60 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m2-ultra",
      "gpuName": "Apple M2 Ultra",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 71.43,
      "decodeLowTokS": 46.43,
      "decodeHighTokS": 96.43,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 60,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.2,
      "localCostPer1MUsd": 0.035,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (60 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (60 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m2-ultra",
      "gpuName": "Apple M2 Ultra",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 40.15,
      "decodeLowTokS": 26.1,
      "decodeHighTokS": 54.2,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 60,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.67,
      "localCostPer1MUsd": 0.06227,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (60 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (60 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m2-ultra",
      "gpuName": "Apple M2 Ultra",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 118.17,
      "decodeLowTokS": 76.81,
      "decodeHighTokS": 159.53,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 60,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 2,
      "localCostPer1MUsd": 0.02116,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (60 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (60 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m2-ultra",
      "gpuName": "Apple M2 Ultra",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 63.82,
      "decodeLowTokS": 41.49,
      "decodeHighTokS": 86.16,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 60,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.1,
      "localCostPer1MUsd": 0.03917,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (60 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (60 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m3",
      "gpuName": "Apple M3",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 9.91,
      "decodeLowTokS": 6.44,
      "decodeHighTokS": 13.38,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 22,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.45,
      "localCostPer1MUsd": 0.09251,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (22 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (22 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m3",
      "gpuName": "Apple M3",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 10.99,
      "decodeLowTokS": 7.14,
      "decodeHighTokS": 14.83,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 22,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.5,
      "localCostPer1MUsd": 0.08342,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (22 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (22 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m3",
      "gpuName": "Apple M3",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 5.61,
      "decodeLowTokS": 3.65,
      "decodeHighTokS": 7.57,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 22,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.25,
      "localCostPer1MUsd": 0.16341,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (22 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (22 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m3",
      "gpuName": "Apple M3",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 9.58,
      "decodeLowTokS": 6.23,
      "decodeHighTokS": 12.94,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 22,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.44,
      "localCostPer1MUsd": 0.09565,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (22 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (22 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m3-ultra",
      "gpuName": "Apple M3 Ultra",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 66.86,
      "decodeLowTokS": 43.46,
      "decodeHighTokS": 90.27,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 60,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.1,
      "localCostPer1MUsd": 0.03739,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (60 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (60 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m3-ultra",
      "gpuName": "Apple M3 Ultra",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 72.75,
      "decodeLowTokS": 47.29,
      "decodeHighTokS": 98.22,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 60,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.2,
      "localCostPer1MUsd": 0.03436,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (60 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (60 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m3-ultra",
      "gpuName": "Apple M3 Ultra",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 40.98,
      "decodeLowTokS": 26.64,
      "decodeHighTokS": 55.33,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 60,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.68,
      "localCostPer1MUsd": 0.061,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (60 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (60 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m3-ultra",
      "gpuName": "Apple M3 Ultra",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 119.97,
      "decodeLowTokS": 77.98,
      "decodeHighTokS": 161.96,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 60,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 2,
      "localCostPer1MUsd": 0.02084,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (60 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (60 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m3-ultra",
      "gpuName": "Apple M3 Ultra",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 65.04,
      "decodeLowTokS": 42.28,
      "decodeHighTokS": 87.81,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 60,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.1,
      "localCostPer1MUsd": 0.03844,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (60 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (60 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m4",
      "gpuName": "Apple M4",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 11.82,
      "decodeLowTokS": 7.68,
      "decodeHighTokS": 15.96,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 22,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.54,
      "localCostPer1MUsd": 0.07755,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (22 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (22 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m4",
      "gpuName": "Apple M4",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 13.1,
      "decodeLowTokS": 8.52,
      "decodeHighTokS": 17.69,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 22,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.6,
      "localCostPer1MUsd": 0.06997,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (22 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (22 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m4",
      "gpuName": "Apple M4",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 6.71,
      "decodeLowTokS": 4.36,
      "decodeHighTokS": 9.06,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 22,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.3,
      "localCostPer1MUsd": 0.13664,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (22 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (22 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m4",
      "gpuName": "Apple M4",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 25.37,
      "decodeLowTokS": 16.49,
      "decodeHighTokS": 34.25,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 22,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.2,
      "localCostPer1MUsd": 0.03613,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (22 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (22 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m4",
      "gpuName": "Apple M4",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 11.43,
      "decodeLowTokS": 7.43,
      "decodeHighTokS": 15.44,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 22,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.52,
      "localCostPer1MUsd": 0.08017,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (22 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (22 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m5",
      "gpuName": "Apple M5",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 14.93,
      "decodeLowTokS": 9.7,
      "decodeHighTokS": 20.15,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 23,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.65,
      "localCostPer1MUsd": 0.06421,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (23 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (23 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m5",
      "gpuName": "Apple M5",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 16.52,
      "decodeLowTokS": 10.74,
      "decodeHighTokS": 22.31,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 23,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.72,
      "localCostPer1MUsd": 0.058,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (23 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (23 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m5",
      "gpuName": "Apple M5",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 8.51,
      "decodeLowTokS": 5.53,
      "decodeHighTokS": 11.48,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 23,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.37,
      "localCostPer1MUsd": 0.11266,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (23 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (23 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m5",
      "gpuName": "Apple M5",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 31.69,
      "decodeLowTokS": 20.6,
      "decodeHighTokS": 42.78,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 23,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.4,
      "localCostPer1MUsd": 0.03025,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (23 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (23 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m5",
      "gpuName": "Apple M5",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 14.44,
      "decodeLowTokS": 9.39,
      "decodeHighTokS": 19.5,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 23,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.63,
      "localCostPer1MUsd": 0.06635,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (23 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (23 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m5-pro",
      "gpuName": "Apple M5 Pro",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 28.66,
      "decodeLowTokS": 18.63,
      "decodeHighTokS": 38.69,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 30,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.96,
      "localCostPer1MUsd": 0.04362,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (30 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (30 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m5-pro",
      "gpuName": "Apple M5 Pro",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 31.58,
      "decodeLowTokS": 20.53,
      "decodeHighTokS": 42.63,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 30,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.1,
      "localCostPer1MUsd": 0.03958,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (30 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (30 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m5-pro",
      "gpuName": "Apple M5 Pro",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 16.64,
      "decodeLowTokS": 10.82,
      "decodeHighTokS": 22.47,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 30,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.55,
      "localCostPer1MUsd": 0.07511,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (30 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (30 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m5-pro",
      "gpuName": "Apple M5 Pro",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 58.03,
      "decodeLowTokS": 37.72,
      "decodeHighTokS": 78.34,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 30,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.9,
      "localCostPer1MUsd": 0.02154,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (30 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (30 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m5-pro",
      "gpuName": "Apple M5 Pro",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 27.77,
      "decodeLowTokS": 18.05,
      "decodeHighTokS": 37.49,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 30,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.93,
      "localCostPer1MUsd": 0.04501,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (30 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (30 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m5-max",
      "gpuName": "Apple M5 Max",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 52.78,
      "decodeLowTokS": 34.31,
      "decodeHighTokS": 71.25,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.5,
      "localCostPer1MUsd": 0.02763,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m5-max",
      "gpuName": "Apple M5 Max",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 57.69,
      "decodeLowTokS": 37.5,
      "decodeHighTokS": 77.89,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.6,
      "localCostPer1MUsd": 0.02528,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m5-max",
      "gpuName": "Apple M5 Max",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 31.7,
      "decodeLowTokS": 20.61,
      "decodeHighTokS": 42.8,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.91,
      "localCostPer1MUsd": 0.046,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m5-max",
      "gpuName": "Apple M5 Max",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 98.85,
      "decodeLowTokS": 64.25,
      "decodeHighTokS": 133.44,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 2.8,
      "localCostPer1MUsd": 0.01475,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "apple-m5-max",
      "gpuName": "Apple M5 Max",
      "vendor": "Apple",
      "architecture": "apple-silicon",
      "backend": "metal",
      "unifiedMemory": true,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 51.27,
      "decodeLowTokS": 33.32,
      "decodeHighTokS": 69.21,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 35,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 1.5,
      "localCostPer1MUsd": 0.02845,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "apple-silicon",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (35 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (35 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "arc-b580",
      "gpuName": "Intel Arc B580",
      "vendor": "Intel",
      "architecture": "intel-xe2",
      "backend": "vulkan",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 42.34,
      "decodeLowTokS": 27.52,
      "decodeHighTokS": 57.17,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 190,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.22,
      "localCostPer1MUsd": 0.18696,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "intel-xe2",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (190 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (190 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "arc-b580",
      "gpuName": "Intel Arc B580",
      "vendor": "Intel",
      "architecture": "intel-xe2",
      "backend": "vulkan",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 46.23,
      "decodeLowTokS": 30.05,
      "decodeHighTokS": 62.42,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 190,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.24,
      "localCostPer1MUsd": 0.17123,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "intel-xe2",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (190 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (190 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "arc-b580",
      "gpuName": "Intel Arc B580",
      "vendor": "Intel",
      "architecture": "intel-xe2",
      "backend": "vulkan",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 25.57,
      "decodeLowTokS": 16.62,
      "decodeHighTokS": 34.52,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 190,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.13,
      "localCostPer1MUsd": 0.30964,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "intel-xe2",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (190 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (190 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "arc-b580",
      "gpuName": "Intel Arc B580",
      "vendor": "Intel",
      "architecture": "intel-xe2",
      "backend": "vulkan",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 41.15,
      "decodeLowTokS": 26.75,
      "decodeHighTokS": 55.55,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 190,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.22,
      "localCostPer1MUsd": 0.19239,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "intel-xe2",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (190 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (190 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "arc-b570",
      "gpuName": "Intel Arc B570",
      "vendor": "Intel",
      "architecture": "intel-xe2",
      "backend": "vulkan",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 36.31,
      "decodeLowTokS": 23.6,
      "decodeHighTokS": 49.02,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 150,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.24,
      "localCostPer1MUsd": 0.17212,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "intel-xe2",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (150 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (150 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "arc-b570",
      "gpuName": "Intel Arc B570",
      "vendor": "Intel",
      "architecture": "intel-xe2",
      "backend": "vulkan",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 39.75,
      "decodeLowTokS": 25.84,
      "decodeHighTokS": 53.67,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 150,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.27,
      "localCostPer1MUsd": 0.15722,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "intel-xe2",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (150 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (150 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "arc-b570",
      "gpuName": "Intel Arc B570",
      "vendor": "Intel",
      "architecture": "intel-xe2",
      "backend": "vulkan",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 35.26,
      "decodeLowTokS": 22.92,
      "decodeHighTokS": 47.6,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 150,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.24,
      "localCostPer1MUsd": 0.17727,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "intel-xe2",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (150 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (150 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-dgx-spark",
      "gpuName": "NVIDIA DGX Spark",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": true,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 39.41,
      "decodeLowTokS": 25.62,
      "decodeHighTokS": 53.2,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 150,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.26,
      "localCostPer1MUsd": 0.15859,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (150 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (150 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-dgx-spark",
      "gpuName": "NVIDIA DGX Spark",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": true,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 43.42,
      "decodeLowTokS": 28.22,
      "decodeHighTokS": 58.61,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 150,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.29,
      "localCostPer1MUsd": 0.14395,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (150 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (150 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-dgx-spark",
      "gpuName": "NVIDIA DGX Spark",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": true,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 22.91,
      "decodeLowTokS": 14.89,
      "decodeHighTokS": 30.93,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 150,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.15,
      "localCostPer1MUsd": 0.27278,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (150 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (150 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-dgx-spark",
      "gpuName": "NVIDIA DGX Spark",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": true,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 79.57,
      "decodeLowTokS": 51.72,
      "decodeHighTokS": 107.42,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 150,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.53,
      "localCostPer1MUsd": 0.07854,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (150 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (150 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-dgx-spark",
      "gpuName": "NVIDIA DGX Spark",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": true,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 38.19,
      "decodeLowTokS": 24.82,
      "decodeHighTokS": 51.56,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 150,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.25,
      "localCostPer1MUsd": 0.16365,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (150 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (150 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "amd-ryzen-ai-max-plus-395",
      "gpuName": "AMD Ryzen AI Max+ 395",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": true,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 27.97,
      "decodeLowTokS": 18.18,
      "decodeHighTokS": 37.76,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 120,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.23,
      "localCostPer1MUsd": 0.17876,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (120 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (120 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "amd-ryzen-ai-max-plus-395",
      "gpuName": "AMD Ryzen AI Max+ 395",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": true,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 30.78,
      "decodeLowTokS": 20.01,
      "decodeHighTokS": 41.56,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 120,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.26,
      "localCostPer1MUsd": 0.16243,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (120 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (120 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "amd-ryzen-ai-max-plus-395",
      "gpuName": "AMD Ryzen AI Max+ 395",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": true,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 16.33,
      "decodeLowTokS": 10.62,
      "decodeHighTokS": 22.05,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 120,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.14,
      "localCostPer1MUsd": 0.30616,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (120 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (120 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "amd-ryzen-ai-max-plus-395",
      "gpuName": "AMD Ryzen AI Max+ 395",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": true,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 55.89,
      "decodeLowTokS": 36.33,
      "decodeHighTokS": 75.45,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 120,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.47,
      "localCostPer1MUsd": 0.08947,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (120 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (120 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "amd-ryzen-ai-max-plus-395",
      "gpuName": "AMD Ryzen AI Max+ 395",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": true,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 27.11,
      "decodeLowTokS": 17.62,
      "decodeHighTokS": 36.6,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 120,
      "wattsBasis": "package-spec",
      "tokensPerWatt": 0.23,
      "localCostPer1MUsd": 0.18441,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "package-spec"
        }
      },
      "caveats": [
        "Power: spec SoC package power (120 W).",
        "This figure divides by SoC package power (CPU, GPU and memory controller on one die), not by discrete-card board power. A unified-memory system and a graphics card are not measuring the same envelope, and neither is the wall draw of the whole machine.",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec SoC package power (120 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-6000-ada",
      "gpuName": "NVIDIA RTX 6000 Ada Generation",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 120.82,
      "decodeLowTokS": 82.73,
      "decodeHighTokS": 158.92,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.4,
      "localCostPer1MUsd": 0.10346,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-6000-ada",
      "gpuName": "NVIDIA RTX 6000 Ada Generation",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 130.45,
      "decodeLowTokS": 89.32,
      "decodeHighTokS": 171.58,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.43,
      "localCostPer1MUsd": 0.09583,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-6000-ada",
      "gpuName": "NVIDIA RTX 6000 Ada Generation",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 76.68,
      "decodeLowTokS": 52.5,
      "decodeHighTokS": 100.86,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.26,
      "localCostPer1MUsd": 0.16301,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-6000-ada",
      "gpuName": "NVIDIA RTX 6000 Ada Generation",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 202.54,
      "decodeLowTokS": 138.68,
      "decodeHighTokS": 266.4,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.68,
      "localCostPer1MUsd": 0.06172,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-6000-ada",
      "gpuName": "NVIDIA RTX 6000 Ada Generation",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 117.82,
      "decodeLowTokS": 80.67,
      "decodeHighTokS": 154.96,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.39,
      "localCostPer1MUsd": 0.1061,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-pro-6000-blackwell",
      "gpuName": "NVIDIA RTX PRO 6000 Blackwell",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 173.33,
      "decodeLowTokS": 112.66,
      "decodeHighTokS": 233.99,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 600,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.29,
      "localCostPer1MUsd": 0.14424,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (600 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (600 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-pro-6000-blackwell",
      "gpuName": "NVIDIA RTX PRO 6000 Blackwell",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 184.75,
      "decodeLowTokS": 120.09,
      "decodeHighTokS": 249.42,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 600,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.31,
      "localCostPer1MUsd": 0.13532,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (600 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (600 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-pro-6000-blackwell",
      "gpuName": "NVIDIA RTX PRO 6000 Blackwell",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 116.92,
      "decodeLowTokS": 76,
      "decodeHighTokS": 157.84,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 600,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.19,
      "localCostPer1MUsd": 0.21382,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (600 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (600 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-pro-6000-blackwell",
      "gpuName": "NVIDIA RTX PRO 6000 Blackwell",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 261.89,
      "decodeLowTokS": 170.23,
      "decodeHighTokS": 353.55,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 600,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.44,
      "localCostPer1MUsd": 0.09546,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (600 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (600 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-pro-6000-blackwell",
      "gpuName": "NVIDIA RTX PRO 6000 Blackwell",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 169.7,
      "decodeLowTokS": 110.31,
      "decodeHighTokS": 229.1,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 600,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.28,
      "localCostPer1MUsd": 0.14732,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (600 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (600 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-l40s",
      "gpuName": "NVIDIA L40S",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 111.77,
      "decodeLowTokS": 76.53,
      "decodeHighTokS": 147.01,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 350,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.32,
      "localCostPer1MUsd": 0.13047,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (350 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (350 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-l40s",
      "gpuName": "NVIDIA L40S",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 120.94,
      "decodeLowTokS": 82.81,
      "decodeHighTokS": 159.08,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 350,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.35,
      "localCostPer1MUsd": 0.12058,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (350 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (350 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-l40s",
      "gpuName": "NVIDIA L40S",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 70.22,
      "decodeLowTokS": 48.08,
      "decodeHighTokS": 92.36,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 350,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.2,
      "localCostPer1MUsd": 0.20767,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (350 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (350 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-l40s",
      "gpuName": "NVIDIA L40S",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 190.97,
      "decodeLowTokS": 130.76,
      "decodeHighTokS": 251.19,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 350,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.55,
      "localCostPer1MUsd": 0.07636,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (350 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (350 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-l40s",
      "gpuName": "NVIDIA L40S",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 108.92,
      "decodeLowTokS": 74.57,
      "decodeHighTokS": 143.26,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 350,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.31,
      "localCostPer1MUsd": 0.13389,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (350 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (350 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-a100-80gb",
      "gpuName": "NVIDIA A100 80GB (PCIe)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 149.19,
      "decodeLowTokS": 102.15,
      "decodeHighTokS": 196.23,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.5,
      "localCostPer1MUsd": 0.08379,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-a100-80gb",
      "gpuName": "NVIDIA A100 80GB (PCIe)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 159.96,
      "decodeLowTokS": 109.52,
      "decodeHighTokS": 210.39,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.53,
      "localCostPer1MUsd": 0.07815,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-a100-80gb",
      "gpuName": "NVIDIA A100 80GB (PCIe)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 97.81,
      "decodeLowTokS": 66.97,
      "decodeHighTokS": 128.65,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.33,
      "localCostPer1MUsd": 0.1278,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-a100-80gb",
      "gpuName": "NVIDIA A100 80GB (PCIe)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 236.13,
      "decodeLowTokS": 161.68,
      "decodeHighTokS": 310.58,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.79,
      "localCostPer1MUsd": 0.05294,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-a100-80gb",
      "gpuName": "NVIDIA A100 80GB (PCIe)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 145.79,
      "decodeLowTokS": 99.83,
      "decodeHighTokS": 191.76,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.49,
      "localCostPer1MUsd": 0.08574,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-h100-80gb",
      "gpuName": "NVIDIA H100 80GB (PCIe)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-hopper",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 185.08,
      "decodeLowTokS": 120.3,
      "decodeHighTokS": 249.86,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 350,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.53,
      "localCostPer1MUsd": 0.07879,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-hopper",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (350 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (350 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-h100-80gb",
      "gpuName": "NVIDIA H100 80GB (PCIe)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-hopper",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 196.72,
      "decodeLowTokS": 127.87,
      "decodeHighTokS": 265.58,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 350,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.56,
      "localCostPer1MUsd": 0.07413,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-hopper",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (350 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (350 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-h100-80gb",
      "gpuName": "NVIDIA H100 80GB (PCIe)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-hopper",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 126.63,
      "decodeLowTokS": 82.31,
      "decodeHighTokS": 170.95,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 350,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.36,
      "localCostPer1MUsd": 0.11516,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-hopper",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (350 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (350 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-h100-80gb",
      "gpuName": "NVIDIA H100 80GB (PCIe)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-hopper",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 273.61,
      "decodeLowTokS": 177.85,
      "decodeHighTokS": 369.37,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 350,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.78,
      "localCostPer1MUsd": 0.0533,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-hopper",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (350 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (350 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-h100-80gb",
      "gpuName": "NVIDIA H100 80GB (PCIe)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-hopper",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 181.37,
      "decodeLowTokS": 117.89,
      "decodeHighTokS": 244.85,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 350,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.52,
      "localCostPer1MUsd": 0.08041,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-hopper",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (350 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (350 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3060-ti",
      "gpuName": "NVIDIA GeForce RTX 3060 Ti",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 46.52,
      "decodeLowTokS": 31.85,
      "decodeHighTokS": 61.19,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 200,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.23,
      "localCostPer1MUsd": 0.17914,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (200 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (200 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3060-ti",
      "gpuName": "NVIDIA GeForce RTX 3060 Ti",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 51.16,
      "decodeLowTokS": 35.03,
      "decodeHighTokS": 67.29,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 200,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.26,
      "localCostPer1MUsd": 0.16289,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (200 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (200 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-3060-ti",
      "gpuName": "NVIDIA GeForce RTX 3060 Ti",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 45.11,
      "decodeLowTokS": 30.88,
      "decodeHighTokS": 59.33,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 200,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.23,
      "localCostPer1MUsd": 0.18475,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (200 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (200 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-pro-4500-blackwell",
      "gpuName": "NVIDIA RTX PRO 4500 Blackwell (Workstation Edition)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 107.61,
      "decodeLowTokS": 69.95,
      "decodeHighTokS": 145.27,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 200,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.54,
      "localCostPer1MUsd": 0.07744,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (200 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (200 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-pro-4500-blackwell",
      "gpuName": "NVIDIA RTX PRO 4500 Blackwell (Workstation Edition)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 116.56,
      "decodeLowTokS": 75.76,
      "decodeHighTokS": 157.36,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 200,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.58,
      "localCostPer1MUsd": 0.07149,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (200 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (200 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-pro-4500-blackwell",
      "gpuName": "NVIDIA RTX PRO 4500 Blackwell (Workstation Edition)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 67.3,
      "decodeLowTokS": 43.74,
      "decodeHighTokS": 90.85,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 200,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.34,
      "localCostPer1MUsd": 0.12383,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (200 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (200 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-pro-4500-blackwell",
      "gpuName": "NVIDIA RTX PRO 4500 Blackwell (Workstation Edition)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 185.5,
      "decodeLowTokS": 120.58,
      "decodeHighTokS": 250.43,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 200,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.93,
      "localCostPer1MUsd": 0.04492,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (200 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (200 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-pro-4500-blackwell",
      "gpuName": "NVIDIA RTX PRO 4500 Blackwell (Workstation Edition)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 104.83,
      "decodeLowTokS": 68.14,
      "decodeHighTokS": 141.52,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 200,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.52,
      "localCostPer1MUsd": 0.0795,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (200 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (200 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5050",
      "gpuName": "NVIDIA GeForce RTX 5050",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 45.5,
      "decodeLowTokS": 29.58,
      "decodeHighTokS": 61.43,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 130,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.35,
      "localCostPer1MUsd": 0.11904,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (130 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (130 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5050",
      "gpuName": "NVIDIA GeForce RTX 5050",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 50.05,
      "decodeLowTokS": 32.53,
      "decodeHighTokS": 67.57,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 130,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.39,
      "localCostPer1MUsd": 0.10822,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (130 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (130 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5050",
      "gpuName": "NVIDIA GeForce RTX 5050",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 44.12,
      "decodeLowTokS": 28.68,
      "decodeHighTokS": 59.56,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 130,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.34,
      "localCostPer1MUsd": 0.12278,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (130 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (130 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-7600",
      "gpuName": "AMD Radeon RX 7600",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 31.09,
      "decodeLowTokS": 20.21,
      "decodeHighTokS": 41.97,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 165,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.19,
      "localCostPer1MUsd": 0.22116,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (165 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (165 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-7600",
      "gpuName": "AMD Radeon RX 7600",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 34.17,
      "decodeLowTokS": 22.21,
      "decodeHighTokS": 46.13,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 165,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.21,
      "localCostPer1MUsd": 0.2012,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (165 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (165 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-7600",
      "gpuName": "AMD Radeon RX 7600",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 30.15,
      "decodeLowTokS": 19.59,
      "decodeHighTokS": 40.7,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 165,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.18,
      "localCostPer1MUsd": 0.22806,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (165 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (165 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-7600-xt",
      "gpuName": "AMD Radeon RX 7600 XT",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 31.09,
      "decodeLowTokS": 20.21,
      "decodeHighTokS": 41.97,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 190,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.16,
      "localCostPer1MUsd": 0.25467,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (190 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (190 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-7600-xt",
      "gpuName": "AMD Radeon RX 7600 XT",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 34.17,
      "decodeLowTokS": 22.21,
      "decodeHighTokS": 46.13,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 190,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.18,
      "localCostPer1MUsd": 0.23169,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (190 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (190 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-7600-xt",
      "gpuName": "AMD Radeon RX 7600 XT",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 18.24,
      "decodeLowTokS": 11.86,
      "decodeHighTokS": 24.63,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 190,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.096,
      "localCostPer1MUsd": 0.43397,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (190 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (190 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-7600-xt",
      "gpuName": "AMD Radeon RX 7600 XT",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 30.15,
      "decodeLowTokS": 19.59,
      "decodeHighTokS": 40.7,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 190,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.16,
      "localCostPer1MUsd": 0.26262,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (190 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (190 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-7900-gre",
      "gpuName": "AMD Radeon RX 7900 GRE",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 56.07,
      "decodeLowTokS": 36.45,
      "decodeHighTokS": 75.7,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 260,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.22,
      "localCostPer1MUsd": 0.19321,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (260 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (260 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-7900-gre",
      "gpuName": "AMD Radeon RX 7900 GRE",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 61.04,
      "decodeLowTokS": 39.68,
      "decodeHighTokS": 82.4,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 260,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.23,
      "localCostPer1MUsd": 0.17748,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (260 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (260 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-7900-gre",
      "gpuName": "AMD Radeon RX 7900 GRE",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 34.3,
      "decodeLowTokS": 22.29,
      "decodeHighTokS": 46.3,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 260,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.13,
      "localCostPer1MUsd": 0.31589,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (260 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (260 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-7900-gre",
      "gpuName": "AMD Radeon RX 7900 GRE",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 54.54,
      "decodeLowTokS": 35.45,
      "decodeHighTokS": 73.62,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 260,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.21,
      "localCostPer1MUsd": 0.19864,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (260 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (260 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-9070-gre",
      "gpuName": "AMD Radeon RX 9070 GRE",
      "vendor": "AMD",
      "architecture": "amd-rdna4",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 47.06,
      "decodeLowTokS": 30.59,
      "decodeHighTokS": 63.54,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 220,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.21,
      "localCostPer1MUsd": 0.19477,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna4",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (220 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (220 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-9070-gre",
      "gpuName": "AMD Radeon RX 9070 GRE",
      "vendor": "AMD",
      "architecture": "amd-rdna4",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 51.41,
      "decodeLowTokS": 33.42,
      "decodeHighTokS": 69.41,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 220,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.23,
      "localCostPer1MUsd": 0.17829,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna4",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (220 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (220 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-9070-gre",
      "gpuName": "AMD Radeon RX 9070 GRE",
      "vendor": "AMD",
      "architecture": "amd-rdna4",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 28.35,
      "decodeLowTokS": 18.43,
      "decodeHighTokS": 38.28,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 220,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.13,
      "localCostPer1MUsd": 0.32329,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna4",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (220 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (220 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rx-9070-gre",
      "gpuName": "AMD Radeon RX 9070 GRE",
      "vendor": "AMD",
      "architecture": "amd-rdna4",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 45.73,
      "decodeLowTokS": 29.72,
      "decodeHighTokS": 61.73,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 220,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.21,
      "localCostPer1MUsd": 0.20046,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna4",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (220 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (220 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-pro-5000-blackwell-48gb",
      "gpuName": "NVIDIA RTX PRO 5000 Blackwell (48 GB)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 144.01,
      "decodeLowTokS": 93.61,
      "decodeHighTokS": 194.42,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.48,
      "localCostPer1MUsd": 0.0868,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-pro-5000-blackwell-48gb",
      "gpuName": "NVIDIA RTX PRO 5000 Blackwell (48 GB)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 154.6,
      "decodeLowTokS": 100.49,
      "decodeHighTokS": 208.72,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.52,
      "localCostPer1MUsd": 0.08085,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-pro-5000-blackwell-48gb",
      "gpuName": "NVIDIA RTX PRO 5000 Blackwell (48 GB)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 93.85,
      "decodeLowTokS": 61,
      "decodeHighTokS": 126.7,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.31,
      "localCostPer1MUsd": 0.13319,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-pro-5000-blackwell-48gb",
      "gpuName": "NVIDIA RTX PRO 5000 Blackwell (48 GB)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 230.28,
      "decodeLowTokS": 149.68,
      "decodeHighTokS": 310.88,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.77,
      "localCostPer1MUsd": 0.05428,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-pro-5000-blackwell-48gb",
      "gpuName": "NVIDIA RTX PRO 5000 Blackwell (48 GB)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 140.68,
      "decodeLowTokS": 91.44,
      "decodeHighTokS": 189.92,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.47,
      "localCostPer1MUsd": 0.08885,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-pro-5000-blackwell-72gb",
      "gpuName": "NVIDIA RTX PRO 5000 Blackwell (72 GB)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 144.01,
      "decodeLowTokS": 93.61,
      "decodeHighTokS": 194.42,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.48,
      "localCostPer1MUsd": 0.0868,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-pro-5000-blackwell-72gb",
      "gpuName": "NVIDIA RTX PRO 5000 Blackwell (72 GB)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 154.6,
      "decodeLowTokS": 100.49,
      "decodeHighTokS": 208.72,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.52,
      "localCostPer1MUsd": 0.08085,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-pro-5000-blackwell-72gb",
      "gpuName": "NVIDIA RTX PRO 5000 Blackwell (72 GB)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 93.85,
      "decodeLowTokS": 61,
      "decodeHighTokS": 126.7,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.31,
      "localCostPer1MUsd": 0.13319,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-pro-5000-blackwell-72gb",
      "gpuName": "NVIDIA RTX PRO 5000 Blackwell (72 GB)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 230.28,
      "decodeLowTokS": 149.68,
      "decodeHighTokS": 310.88,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.77,
      "localCostPer1MUsd": 0.05428,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-pro-5000-blackwell-72gb",
      "gpuName": "NVIDIA RTX PRO 5000 Blackwell (72 GB)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 140.68,
      "decodeLowTokS": 91.44,
      "decodeHighTokS": 189.92,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.47,
      "localCostPer1MUsd": 0.08885,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-pro-4000-blackwell",
      "gpuName": "NVIDIA RTX PRO 4000 Blackwell",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 85.9,
      "decodeLowTokS": 55.83,
      "decodeHighTokS": 115.96,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": null,
      "wattsBasis": null,
      "tokensPerWatt": null,
      "localCostPer1MUsd": null,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": null,
        "tokensPerWattTier": null,
        "localCostBadge": null,
        "localCostTier": null,
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "none"
        }
      },
      "caveats": []
    },
    {
      "gpuId": "rtx-pro-4000-blackwell",
      "gpuName": "NVIDIA RTX PRO 4000 Blackwell",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 93.54,
      "decodeLowTokS": 60.8,
      "decodeHighTokS": 126.28,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": null,
      "wattsBasis": null,
      "tokensPerWatt": null,
      "localCostPer1MUsd": null,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": null,
        "tokensPerWattTier": null,
        "localCostBadge": null,
        "localCostTier": null,
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "none"
        }
      },
      "caveats": []
    },
    {
      "gpuId": "rtx-pro-4000-blackwell",
      "gpuName": "NVIDIA RTX PRO 4000 Blackwell",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 52.45,
      "decodeLowTokS": 34.1,
      "decodeHighTokS": 70.81,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": null,
      "wattsBasis": null,
      "tokensPerWatt": null,
      "localCostPer1MUsd": null,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": null,
        "tokensPerWattTier": null,
        "localCostBadge": null,
        "localCostTier": null,
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "none"
        }
      },
      "caveats": []
    },
    {
      "gpuId": "rtx-pro-4000-blackwell",
      "gpuName": "NVIDIA RTX PRO 4000 Blackwell",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 155.3,
      "decodeLowTokS": 100.95,
      "decodeHighTokS": 209.66,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": null,
      "wattsBasis": null,
      "tokensPerWatt": null,
      "localCostPer1MUsd": null,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": null,
        "tokensPerWattTier": null,
        "localCostBadge": null,
        "localCostTier": null,
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "none"
        }
      },
      "caveats": []
    },
    {
      "gpuId": "rtx-pro-4000-blackwell",
      "gpuName": "NVIDIA RTX PRO 4000 Blackwell",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 83.54,
      "decodeLowTokS": 54.3,
      "decodeHighTokS": 112.78,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": null,
      "wattsBasis": null,
      "tokensPerWatt": null,
      "localCostPer1MUsd": null,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": null,
        "tokensPerWattTier": null,
        "localCostBadge": null,
        "localCostTier": null,
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "none"
        }
      },
      "caveats": []
    },
    {
      "gpuId": "rtx-pro-2000-blackwell",
      "gpuName": "NVIDIA RTX PRO 2000 Blackwell",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 41.37,
      "decodeLowTokS": 26.89,
      "decodeHighTokS": 55.86,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 70,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.59,
      "localCostPer1MUsd": 0.07049,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (70 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (70 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-pro-2000-blackwell",
      "gpuName": "NVIDIA RTX PRO 2000 Blackwell",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 45.56,
      "decodeLowTokS": 29.61,
      "decodeHighTokS": 61.51,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 70,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.65,
      "localCostPer1MUsd": 0.06402,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (70 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (70 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-pro-2000-blackwell",
      "gpuName": "NVIDIA RTX PRO 2000 Blackwell",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 24.1,
      "decodeLowTokS": 15.67,
      "decodeHighTokS": 32.54,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 70,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.34,
      "localCostPer1MUsd": 0.12101,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (70 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (70 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-pro-2000-blackwell",
      "gpuName": "NVIDIA RTX PRO 2000 Blackwell",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 40.1,
      "decodeLowTokS": 26.07,
      "decodeHighTokS": 54.14,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 70,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.57,
      "localCostPer1MUsd": 0.07273,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (70 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (70 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-pro-6000-blackwell-max-q",
      "gpuName": "NVIDIA RTX PRO 6000 Blackwell (Max-Q Edition)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 173.33,
      "decodeLowTokS": 112.66,
      "decodeHighTokS": 233.99,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": null,
      "wattsBasis": null,
      "tokensPerWatt": null,
      "localCostPer1MUsd": null,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": null,
        "tokensPerWattTier": null,
        "localCostBadge": null,
        "localCostTier": null,
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "none"
        }
      },
      "caveats": []
    },
    {
      "gpuId": "rtx-pro-6000-blackwell-max-q",
      "gpuName": "NVIDIA RTX PRO 6000 Blackwell (Max-Q Edition)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 184.75,
      "decodeLowTokS": 120.09,
      "decodeHighTokS": 249.42,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": null,
      "wattsBasis": null,
      "tokensPerWatt": null,
      "localCostPer1MUsd": null,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": null,
        "tokensPerWattTier": null,
        "localCostBadge": null,
        "localCostTier": null,
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "none"
        }
      },
      "caveats": []
    },
    {
      "gpuId": "rtx-pro-6000-blackwell-max-q",
      "gpuName": "NVIDIA RTX PRO 6000 Blackwell (Max-Q Edition)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 116.92,
      "decodeLowTokS": 76,
      "decodeHighTokS": 157.84,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": null,
      "wattsBasis": null,
      "tokensPerWatt": null,
      "localCostPer1MUsd": null,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": null,
        "tokensPerWattTier": null,
        "localCostBadge": null,
        "localCostTier": null,
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "none"
        }
      },
      "caveats": []
    },
    {
      "gpuId": "rtx-pro-6000-blackwell-max-q",
      "gpuName": "NVIDIA RTX PRO 6000 Blackwell (Max-Q Edition)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 261.89,
      "decodeLowTokS": 170.23,
      "decodeHighTokS": 353.55,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": null,
      "wattsBasis": null,
      "tokensPerWatt": null,
      "localCostPer1MUsd": null,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": null,
        "tokensPerWattTier": null,
        "localCostBadge": null,
        "localCostTier": null,
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "none"
        }
      },
      "caveats": []
    },
    {
      "gpuId": "rtx-pro-6000-blackwell-max-q",
      "gpuName": "NVIDIA RTX PRO 6000 Blackwell (Max-Q Edition)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 169.7,
      "decodeLowTokS": 110.31,
      "decodeHighTokS": 229.1,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": null,
      "wattsBasis": null,
      "tokensPerWatt": null,
      "localCostPer1MUsd": null,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": null,
        "tokensPerWattTier": null,
        "localCostBadge": null,
        "localCostTier": null,
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "none"
        }
      },
      "caveats": []
    },
    {
      "gpuId": "rtx-5000-ada",
      "gpuName": "NVIDIA RTX 5000 Ada Generation",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 81.32,
      "decodeLowTokS": 55.68,
      "decodeHighTokS": 106.96,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 250,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.33,
      "localCostPer1MUsd": 0.1281,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (250 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (250 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5000-ada",
      "gpuName": "NVIDIA RTX 5000 Ada Generation",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 88.66,
      "decodeLowTokS": 60.7,
      "decodeHighTokS": 116.61,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 250,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.35,
      "localCostPer1MUsd": 0.11749,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (250 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (250 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5000-ada",
      "gpuName": "NVIDIA RTX 5000 Ada Generation",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 49.41,
      "decodeLowTokS": 33.83,
      "decodeHighTokS": 64.99,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 250,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.2,
      "localCostPer1MUsd": 0.21081,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (250 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (250 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5000-ada",
      "gpuName": "NVIDIA RTX 5000 Ada Generation",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 148.56,
      "decodeLowTokS": 101.72,
      "decodeHighTokS": 195.4,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 250,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.59,
      "localCostPer1MUsd": 0.07012,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (250 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (250 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5000-ada",
      "gpuName": "NVIDIA RTX 5000 Ada Generation",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 79.06,
      "decodeLowTokS": 54.13,
      "decodeHighTokS": 103.98,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 250,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.32,
      "localCostPer1MUsd": 0.13176,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (250 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (250 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-a6000",
      "gpuName": "NVIDIA RTX A6000",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 74.21,
      "decodeLowTokS": 50.81,
      "decodeHighTokS": 97.61,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.25,
      "localCostPer1MUsd": 0.16844,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-a6000",
      "gpuName": "NVIDIA RTX A6000",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 81.05,
      "decodeLowTokS": 55.49,
      "decodeHighTokS": 106.6,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.27,
      "localCostPer1MUsd": 0.15423,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-a6000",
      "gpuName": "NVIDIA RTX A6000",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 44.75,
      "decodeLowTokS": 30.64,
      "decodeHighTokS": 58.86,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.15,
      "localCostPer1MUsd": 0.27933,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-a6000",
      "gpuName": "NVIDIA RTX A6000",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 137.79,
      "decodeLowTokS": 94.35,
      "decodeHighTokS": 181.24,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.46,
      "localCostPer1MUsd": 0.09072,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-a6000",
      "gpuName": "NVIDIA RTX A6000",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 72.11,
      "decodeLowTokS": 49.37,
      "decodeHighTokS": 94.84,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.24,
      "localCostPer1MUsd": 0.17336,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-a5000",
      "gpuName": "NVIDIA RTX A5000",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 74.21,
      "decodeLowTokS": 50.81,
      "decodeHighTokS": 97.61,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 230,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.32,
      "localCostPer1MUsd": 0.12914,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (230 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (230 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-a5000",
      "gpuName": "NVIDIA RTX A5000",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 81.05,
      "decodeLowTokS": 55.49,
      "decodeHighTokS": 106.6,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 230,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.35,
      "localCostPer1MUsd": 0.11824,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (230 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (230 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-a5000",
      "gpuName": "NVIDIA RTX A5000",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 44.75,
      "decodeLowTokS": 30.64,
      "decodeHighTokS": 58.86,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 230,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.19,
      "localCostPer1MUsd": 0.21415,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (230 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (230 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-a5000",
      "gpuName": "NVIDIA RTX A5000",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 137.79,
      "decodeLowTokS": 94.35,
      "decodeHighTokS": 181.24,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 230,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.6,
      "localCostPer1MUsd": 0.06955,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (230 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (230 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-a5000",
      "gpuName": "NVIDIA RTX A5000",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 72.11,
      "decodeLowTokS": 49.37,
      "decodeHighTokS": 94.84,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 230,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.31,
      "localCostPer1MUsd": 0.13291,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (230 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (230 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "radeon-pro-w7900",
      "gpuName": "AMD Radeon PRO W7900",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 76.59,
      "decodeLowTokS": 49.78,
      "decodeHighTokS": 103.4,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 295,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.26,
      "localCostPer1MUsd": 0.16048,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (295 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (295 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "radeon-pro-w7900",
      "gpuName": "AMD Radeon PRO W7900",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 82.72,
      "decodeLowTokS": 53.77,
      "decodeHighTokS": 111.68,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 295,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.28,
      "localCostPer1MUsd": 0.14859,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (295 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (295 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "radeon-pro-w7900",
      "gpuName": "AMD Radeon PRO W7900",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 48.53,
      "decodeLowTokS": 31.54,
      "decodeHighTokS": 65.52,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 295,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.16,
      "localCostPer1MUsd": 0.25328,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (295 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (295 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "radeon-pro-w7900",
      "gpuName": "AMD Radeon PRO W7900",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 128.79,
      "decodeLowTokS": 83.71,
      "decodeHighTokS": 173.86,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 295,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.44,
      "localCostPer1MUsd": 0.09544,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (295 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (295 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "radeon-pro-w7900",
      "gpuName": "AMD Radeon PRO W7900",
      "vendor": "AMD",
      "architecture": "amd-rdna3",
      "backend": "rocm",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 74.68,
      "decodeLowTokS": 48.54,
      "decodeHighTokS": 100.81,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 295,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.25,
      "localCostPer1MUsd": 0.1646,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "amd-rdna3",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (295 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (295 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "intel-arc-pro-b60",
      "gpuName": "Intel Arc Pro B60",
      "vendor": "Intel",
      "architecture": "intel-xe2",
      "backend": "vulkan",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 42.34,
      "decodeLowTokS": 27.52,
      "decodeHighTokS": 57.17,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 200,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.21,
      "localCostPer1MUsd": 0.1968,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "intel-xe2",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (200 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (200 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "intel-arc-pro-b60",
      "gpuName": "Intel Arc Pro B60",
      "vendor": "Intel",
      "architecture": "intel-xe2",
      "backend": "vulkan",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 46.23,
      "decodeLowTokS": 30.05,
      "decodeHighTokS": 62.42,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 200,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.23,
      "localCostPer1MUsd": 0.18024,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "intel-xe2",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (200 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (200 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "intel-arc-pro-b60",
      "gpuName": "Intel Arc Pro B60",
      "vendor": "Intel",
      "architecture": "intel-xe2",
      "backend": "vulkan",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 25.57,
      "decodeLowTokS": 16.62,
      "decodeHighTokS": 34.52,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 200,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.13,
      "localCostPer1MUsd": 0.32593,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "intel-xe2",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (200 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (200 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "intel-arc-pro-b60",
      "gpuName": "Intel Arc Pro B60",
      "vendor": "Intel",
      "architecture": "intel-xe2",
      "backend": "vulkan",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 78.41,
      "decodeLowTokS": 50.97,
      "decodeHighTokS": 105.85,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 200,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.39,
      "localCostPer1MUsd": 0.10628,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "intel-xe2",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (200 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (200 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "intel-arc-pro-b60",
      "gpuName": "Intel Arc Pro B60",
      "vendor": "Intel",
      "architecture": "intel-xe2",
      "backend": "vulkan",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 41.15,
      "decodeLowTokS": 26.75,
      "decodeHighTokS": 55.55,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 200,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.21,
      "localCostPer1MUsd": 0.20252,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "intel-xe2",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (200 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (200 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "intel-arc-pro-b50",
      "gpuName": "Intel Arc Pro B50",
      "vendor": "Intel",
      "architecture": "intel-xe2",
      "backend": "vulkan",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 22.76,
      "decodeLowTokS": 14.8,
      "decodeHighTokS": 30.73,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 70,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.33,
      "localCostPer1MUsd": 0.12814,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "intel-xe2",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (70 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (70 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "intel-arc-pro-b50",
      "gpuName": "Intel Arc Pro B50",
      "vendor": "Intel",
      "architecture": "intel-xe2",
      "backend": "vulkan",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 25.07,
      "decodeLowTokS": 16.3,
      "decodeHighTokS": 33.84,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 70,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.36,
      "localCostPer1MUsd": 0.11634,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "intel-xe2",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (70 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (70 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "intel-arc-pro-b50",
      "gpuName": "Intel Arc Pro B50",
      "vendor": "Intel",
      "architecture": "intel-xe2",
      "backend": "vulkan",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 13.25,
      "decodeLowTokS": 8.61,
      "decodeHighTokS": 17.89,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 70,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.19,
      "localCostPer1MUsd": 0.22014,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "intel-xe2",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (70 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (70 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "intel-arc-pro-b50",
      "gpuName": "Intel Arc Pro B50",
      "vendor": "Intel",
      "architecture": "intel-xe2",
      "backend": "vulkan",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 22.06,
      "decodeLowTokS": 14.34,
      "decodeHighTokS": 29.78,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 70,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.32,
      "localCostPer1MUsd": 0.13221,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "intel-xe2",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (70 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (70 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-l4",
      "gpuName": "NVIDIA L4",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 46.42,
      "decodeLowTokS": 31.78,
      "decodeHighTokS": 61.05,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 72,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.64,
      "localCostPer1MUsd": 0.06463,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (72 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (72 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-l4",
      "gpuName": "NVIDIA L4",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 51.05,
      "decodeLowTokS": 34.95,
      "decodeHighTokS": 67.14,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 72,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.71,
      "localCostPer1MUsd": 0.05877,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (72 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (72 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-l4",
      "gpuName": "NVIDIA L4",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 27.18,
      "decodeLowTokS": 18.61,
      "decodeHighTokS": 35.75,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 72,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.38,
      "localCostPer1MUsd": 0.11037,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (72 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (72 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-l4",
      "gpuName": "NVIDIA L4",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 92.1,
      "decodeLowTokS": 63.06,
      "decodeHighTokS": 121.14,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 72,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 1.3,
      "localCostPer1MUsd": 0.03257,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (72 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (72 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-l4",
      "gpuName": "NVIDIA L4",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 45,
      "decodeLowTokS": 30.81,
      "decodeHighTokS": 59.19,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 72,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.63,
      "localCostPer1MUsd": 0.06666,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (72 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (72 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-a10",
      "gpuName": "NVIDIA A10",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 60.17,
      "decodeLowTokS": 41.2,
      "decodeHighTokS": 79.14,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 150,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.4,
      "localCostPer1MUsd": 0.10387,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (150 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (150 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-a10",
      "gpuName": "NVIDIA A10",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 65.94,
      "decodeLowTokS": 45.15,
      "decodeHighTokS": 86.74,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 150,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.44,
      "localCostPer1MUsd": 0.09478,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (150 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (150 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-a10",
      "gpuName": "NVIDIA A10",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 35.75,
      "decodeLowTokS": 24.48,
      "decodeHighTokS": 47.02,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 150,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.24,
      "localCostPer1MUsd": 0.17484,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (150 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (150 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-a10",
      "gpuName": "NVIDIA A10",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 115.47,
      "decodeLowTokS": 79.06,
      "decodeHighTokS": 151.87,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 150,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.77,
      "localCostPer1MUsd": 0.05413,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (150 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (150 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-a10",
      "gpuName": "NVIDIA A10",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 58.4,
      "decodeLowTokS": 39.99,
      "decodeHighTokS": 76.82,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 150,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.39,
      "localCostPer1MUsd": 0.10702,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · provisional fit (n=1)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · provisional fit (n=1)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (150 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (150 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-l40",
      "gpuName": "NVIDIA L40",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 111.77,
      "decodeLowTokS": 76.53,
      "decodeHighTokS": 147.01,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.37,
      "localCostPer1MUsd": 0.11184,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-l40",
      "gpuName": "NVIDIA L40",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 120.94,
      "decodeLowTokS": 82.81,
      "decodeHighTokS": 159.08,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.4,
      "localCostPer1MUsd": 0.10335,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-l40",
      "gpuName": "NVIDIA L40",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 70.22,
      "decodeLowTokS": 48.08,
      "decodeHighTokS": 92.36,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.23,
      "localCostPer1MUsd": 0.178,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-l40",
      "gpuName": "NVIDIA L40",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 190.97,
      "decodeLowTokS": 130.76,
      "decodeHighTokS": 251.19,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.64,
      "localCostPer1MUsd": 0.06545,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-l40",
      "gpuName": "NVIDIA L40",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 108.92,
      "decodeLowTokS": 74.57,
      "decodeHighTokS": 143.26,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 300,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.36,
      "localCostPer1MUsd": 0.11477,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · fitted (n=12)",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · fitted (n=12)",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (300 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (300 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-h100-sxm-80gb",
      "gpuName": "NVIDIA H100 80GB (SXM)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-hopper",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 242.09,
      "decodeLowTokS": 157.36,
      "decodeHighTokS": 326.82,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 700,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.35,
      "localCostPer1MUsd": 0.12048,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-hopper",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (700 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (700 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-h100-sxm-80gb",
      "gpuName": "NVIDIA H100 80GB (SXM)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-hopper",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 253.81,
      "decodeLowTokS": 164.98,
      "decodeHighTokS": 342.65,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 700,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.36,
      "localCostPer1MUsd": 0.11491,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-hopper",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (700 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (700 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-h100-sxm-80gb",
      "gpuName": "NVIDIA H100 80GB (SXM)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-hopper",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 177.95,
      "decodeLowTokS": 115.66,
      "decodeHighTokS": 240.23,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 700,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.25,
      "localCostPer1MUsd": 0.16391,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-hopper",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (700 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (700 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-h100-sxm-80gb",
      "gpuName": "NVIDIA H100 80GB (SXM)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-hopper",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 323.93,
      "decodeLowTokS": 210.55,
      "decodeHighTokS": 437.3,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": 700,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.46,
      "localCostPer1MUsd": 0.09004,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-hopper",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (700 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (700 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "nvidia-h100-sxm-80gb",
      "gpuName": "NVIDIA H100 80GB (SXM)",
      "vendor": "NVIDIA",
      "architecture": "nvidia-hopper",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 238.28,
      "decodeLowTokS": 154.88,
      "decodeHighTokS": 321.68,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 700,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.34,
      "localCostPer1MUsd": 0.1224,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-hopper",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (700 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (700 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5080-laptop",
      "gpuName": "NVIDIA GeForce RTX 5080 Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 107.61,
      "decodeLowTokS": 69.95,
      "decodeHighTokS": 145.27,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 150,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.72,
      "localCostPer1MUsd": 0.05808,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (150 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (150 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5080-laptop",
      "gpuName": "NVIDIA GeForce RTX 5080 Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 116.56,
      "decodeLowTokS": 75.76,
      "decodeHighTokS": 157.36,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 150,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.78,
      "localCostPer1MUsd": 0.05362,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (150 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (150 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5080-laptop",
      "gpuName": "NVIDIA GeForce RTX 5080 Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 67.3,
      "decodeLowTokS": 43.74,
      "decodeHighTokS": 90.85,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 150,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.45,
      "localCostPer1MUsd": 0.09287,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (150 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (150 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5080-laptop",
      "gpuName": "NVIDIA GeForce RTX 5080 Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 104.83,
      "decodeLowTokS": 68.14,
      "decodeHighTokS": 141.52,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 150,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.7,
      "localCostPer1MUsd": 0.05962,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (150 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (150 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5070-ti-laptop",
      "gpuName": "NVIDIA GeForce RTX 5070 Ti Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 85.9,
      "decodeLowTokS": 55.83,
      "decodeHighTokS": 115.96,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 115,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.75,
      "localCostPer1MUsd": 0.05578,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (115 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (115 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5070-ti-laptop",
      "gpuName": "NVIDIA GeForce RTX 5070 Ti Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 93.54,
      "decodeLowTokS": 60.8,
      "decodeHighTokS": 126.28,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 115,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.81,
      "localCostPer1MUsd": 0.05122,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (115 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (115 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5070-ti-laptop",
      "gpuName": "NVIDIA GeForce RTX 5070 Ti Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 52.45,
      "decodeLowTokS": 34.1,
      "decodeHighTokS": 70.81,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": 115,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.46,
      "localCostPer1MUsd": 0.09135,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (115 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (115 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5070-ti-laptop",
      "gpuName": "NVIDIA GeForce RTX 5070 Ti Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 83.54,
      "decodeLowTokS": 54.3,
      "decodeHighTokS": 112.78,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 115,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.73,
      "localCostPer1MUsd": 0.05736,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (115 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (115 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5070-laptop",
      "gpuName": "NVIDIA GeForce RTX 5070 Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 53.51,
      "decodeLowTokS": 34.78,
      "decodeHighTokS": 72.24,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": 100,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.54,
      "localCostPer1MUsd": 0.07787,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (100 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (100 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5070-laptop",
      "gpuName": "NVIDIA GeForce RTX 5070 Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 58.74,
      "decodeLowTokS": 38.18,
      "decodeHighTokS": 79.3,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": 100,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.59,
      "localCostPer1MUsd": 0.07093,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (100 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (100 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5070-laptop",
      "gpuName": "NVIDIA GeForce RTX 5070 Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 51.91,
      "decodeLowTokS": 33.74,
      "decodeHighTokS": 70.08,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": 100,
      "wattsBasis": "board-spec",
      "tokensPerWatt": 0.52,
      "localCostPer1MUsd": 0.08027,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": "ESTIMATED · uncalibrated",
        "tokensPerWattTier": "estimated",
        "localCostBadge": "ESTIMATED · uncalibrated",
        "localCostTier": "estimated",
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "estimated",
          "basis": "board-spec"
        }
      },
      "caveats": [
        "Power: spec board power (TGP/TBP) (100 W).",
        "Electricity only. Excludes the hardware itself, amortisation, idle draw, PSU inefficiency and the rest of the system — this is not the cost of owning the machine. See /en/cost for break-even including hardware.",
        "Power: spec board power (TGP/TBP) (100 W). Rate: $0.15/kWh."
      ]
    },
    {
      "gpuId": "rtx-5060-laptop",
      "gpuName": "NVIDIA GeForce RTX 5060 Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 53.51,
      "decodeLowTokS": 34.78,
      "decodeHighTokS": 72.24,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": null,
      "wattsBasis": null,
      "tokensPerWatt": null,
      "localCostPer1MUsd": null,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": null,
        "tokensPerWattTier": null,
        "localCostBadge": null,
        "localCostTier": null,
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "none"
        }
      },
      "caveats": []
    },
    {
      "gpuId": "rtx-5060-laptop",
      "gpuName": "NVIDIA GeForce RTX 5060 Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 58.74,
      "decodeLowTokS": 38.18,
      "decodeHighTokS": 79.3,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": null,
      "wattsBasis": null,
      "tokensPerWatt": null,
      "localCostPer1MUsd": null,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": null,
        "tokensPerWattTier": null,
        "localCostBadge": null,
        "localCostTier": null,
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "none"
        }
      },
      "caveats": []
    },
    {
      "gpuId": "rtx-5060-laptop",
      "gpuName": "NVIDIA GeForce RTX 5060 Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 51.91,
      "decodeLowTokS": 33.74,
      "decodeHighTokS": 70.08,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": null,
      "wattsBasis": null,
      "tokensPerWatt": null,
      "localCostPer1MUsd": null,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": null,
        "tokensPerWattTier": null,
        "localCostBadge": null,
        "localCostTier": null,
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "none"
        }
      },
      "caveats": []
    },
    {
      "gpuId": "rtx-4080-laptop",
      "gpuName": "NVIDIA GeForce RTX 4080 Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 63.91,
      "decodeLowTokS": 43.76,
      "decodeHighTokS": 84.06,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": null,
      "wattsBasis": null,
      "tokensPerWatt": null,
      "localCostPer1MUsd": null,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": null,
        "tokensPerWattTier": null,
        "localCostBadge": null,
        "localCostTier": null,
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "none"
        }
      },
      "caveats": []
    },
    {
      "gpuId": "rtx-4080-laptop",
      "gpuName": "NVIDIA GeForce RTX 4080 Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 69.98,
      "decodeLowTokS": 47.91,
      "decodeHighTokS": 92.04,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": null,
      "wattsBasis": null,
      "tokensPerWatt": null,
      "localCostPer1MUsd": null,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": null,
        "tokensPerWattTier": null,
        "localCostBadge": null,
        "localCostTier": null,
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "none"
        }
      },
      "caveats": []
    },
    {
      "gpuId": "rtx-4080-laptop",
      "gpuName": "NVIDIA GeForce RTX 4080 Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 38.12,
      "decodeLowTokS": 26.1,
      "decodeHighTokS": 50.14,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": null,
      "wattsBasis": null,
      "tokensPerWatt": null,
      "localCostPer1MUsd": null,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": null,
        "tokensPerWattTier": null,
        "localCostBadge": null,
        "localCostTier": null,
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "none"
        }
      },
      "caveats": []
    },
    {
      "gpuId": "rtx-4080-laptop",
      "gpuName": "NVIDIA GeForce RTX 4080 Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ada",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 62.05,
      "decodeLowTokS": 42.48,
      "decodeHighTokS": 81.61,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": null,
      "wattsBasis": null,
      "tokensPerWatt": null,
      "localCostPer1MUsd": null,
      "provenance": {
        "decodeBadge": "ESTIMATED · fitted (n=12)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": null,
        "tokensPerWattTier": null,
        "localCostBadge": null,
        "localCostTier": null,
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "fitted",
          "architecture": "nvidia-ada",
          "fittedFromRuns": 12,
          "mapePct": 7.68
        },
        "power": {
          "tier": "none"
        }
      },
      "caveats": []
    },
    {
      "gpuId": "rtx-3080-ti-laptop",
      "gpuName": "NVIDIA GeForce RTX 3080 Ti Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 52.38,
      "decodeLowTokS": 35.87,
      "decodeHighTokS": 68.9,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": null,
      "wattsBasis": null,
      "tokensPerWatt": null,
      "localCostPer1MUsd": null,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": null,
        "tokensPerWattTier": null,
        "localCostBadge": null,
        "localCostTier": null,
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "none"
        }
      },
      "caveats": []
    },
    {
      "gpuId": "rtx-3080-ti-laptop",
      "gpuName": "NVIDIA GeForce RTX 3080 Ti Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 57.52,
      "decodeLowTokS": 39.39,
      "decodeHighTokS": 75.66,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": null,
      "wattsBasis": null,
      "tokensPerWatt": null,
      "localCostPer1MUsd": null,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": null,
        "tokensPerWattTier": null,
        "localCostBadge": null,
        "localCostTier": null,
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "none"
        }
      },
      "caveats": []
    },
    {
      "gpuId": "rtx-3080-ti-laptop",
      "gpuName": "NVIDIA GeForce RTX 3080 Ti Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 30.87,
      "decodeLowTokS": 21.13,
      "decodeHighTokS": 40.6,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": null,
      "wattsBasis": null,
      "tokensPerWatt": null,
      "localCostPer1MUsd": null,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": null,
        "tokensPerWattTier": null,
        "localCostBadge": null,
        "localCostTier": null,
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "none"
        }
      },
      "caveats": []
    },
    {
      "gpuId": "rtx-3080-ti-laptop",
      "gpuName": "NVIDIA GeForce RTX 3080 Ti Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-ampere",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 50.81,
      "decodeLowTokS": 34.79,
      "decodeHighTokS": 66.83,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": null,
      "wattsBasis": null,
      "tokensPerWatt": null,
      "localCostPer1MUsd": null,
      "provenance": {
        "decodeBadge": "ESTIMATED · provisional fit (n=1)",
        "decodeTier": "estimated",
        "tokensPerWattBadge": null,
        "tokensPerWattTier": null,
        "localCostBadge": null,
        "localCostTier": null,
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "provisional",
          "architecture": "nvidia-ampere",
          "fittedFromRuns": 1,
          "mapePct": null
        },
        "power": {
          "tier": "none"
        }
      },
      "caveats": []
    },
    {
      "gpuId": "rtx-pro-5000-blackwell-laptop",
      "gpuName": "NVIDIA RTX PRO 5000 Blackwell Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "llama-3.1-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 107.61,
      "decodeLowTokS": 69.95,
      "decodeHighTokS": 145.27,
      "requiredVramGb": 6.17,
      "fitStatus": "great",
      "watts": null,
      "wattsBasis": null,
      "tokensPerWatt": null,
      "localCostPer1MUsd": null,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": null,
        "tokensPerWattTier": null,
        "localCostBadge": null,
        "localCostTier": null,
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "none"
        }
      },
      "caveats": []
    },
    {
      "gpuId": "rtx-pro-5000-blackwell-laptop",
      "gpuName": "NVIDIA RTX PRO 5000 Blackwell Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen-2.5-7b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 116.56,
      "decodeLowTokS": 75.76,
      "decodeHighTokS": 157.36,
      "requiredVramGb": 5.62,
      "fitStatus": "great",
      "watts": null,
      "wattsBasis": null,
      "tokensPerWatt": null,
      "localCostPer1MUsd": null,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": null,
        "tokensPerWattTier": null,
        "localCostBadge": null,
        "localCostTier": null,
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "none"
        }
      },
      "caveats": []
    },
    {
      "gpuId": "rtx-pro-5000-blackwell-laptop",
      "gpuName": "NVIDIA RTX PRO 5000 Blackwell Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-14b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 67.3,
      "decodeLowTokS": 43.74,
      "decodeHighTokS": 90.85,
      "requiredVramGb": 10.41,
      "fitStatus": "great",
      "watts": null,
      "wattsBasis": null,
      "tokensPerWatt": null,
      "localCostPer1MUsd": null,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": null,
        "tokensPerWattTier": null,
        "localCostBadge": null,
        "localCostTier": null,
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "none"
        }
      },
      "caveats": []
    },
    {
      "gpuId": "rtx-pro-5000-blackwell-laptop",
      "gpuName": "NVIDIA RTX PRO 5000 Blackwell Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-30b-a3b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 185.5,
      "decodeLowTokS": 120.58,
      "decodeHighTokS": 250.43,
      "requiredVramGb": 19.62,
      "fitStatus": "great",
      "watts": null,
      "wattsBasis": null,
      "tokensPerWatt": null,
      "localCostPer1MUsd": null,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": null,
        "tokensPerWattTier": null,
        "localCostBadge": null,
        "localCostTier": null,
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "none"
        }
      },
      "caveats": []
    },
    {
      "gpuId": "rtx-pro-5000-blackwell-laptop",
      "gpuName": "NVIDIA RTX PRO 5000 Blackwell Laptop GPU",
      "vendor": "NVIDIA",
      "architecture": "nvidia-blackwell",
      "backend": "cuda",
      "unifiedMemory": false,
      "variantId": "qwen3-8b",
      "quant": "Q4_K_M",
      "contextLength": 4096,
      "runtime": "llama.cpp",
      "decodeTokS": 104.83,
      "decodeLowTokS": 68.14,
      "decodeHighTokS": 141.52,
      "requiredVramGb": 6.35,
      "fitStatus": "great",
      "watts": null,
      "wattsBasis": null,
      "tokensPerWatt": null,
      "localCostPer1MUsd": null,
      "provenance": {
        "decodeBadge": "ESTIMATED · uncalibrated",
        "decodeTier": "estimated",
        "tokensPerWattBadge": null,
        "tokensPerWattTier": null,
        "localCostBadge": null,
        "localCostTier": null,
        "mixed": false,
        "offloadWarning": false,
        "speed": {
          "tier": "estimated",
          "calibration": "seed",
          "architecture": "nvidia-blackwell",
          "fittedFromRuns": 0,
          "mapePct": null
        },
        "power": {
          "tier": "none"
        }
      },
      "caveats": []
    }
  ]
}