95 devices · 4 brands
GPU database
Every GPU, laptop chip and Mac we track, with the numbers that decide local LLM performance: VRAM sets which models fit, memory bandwidth sets how fast they generate.
Specs verified 2026-10-07 · Prices verified 2026-09-29
Tracked954 brands
Recommendation of the monthSeptember 2026
NVIDIA GeForce RTX 3090 24 GB (used)
At $600–750 used it is still the cheapest way to 24 GB of VRAM. It runs the same 27–32B models as an RTX 4090 at Q4 for about half the used price ($1,200–1,500), with 936 GB/s of memory bandwidth against the RTX 4090's 1,008 GB/s. Check VRAM health and thermals before you pay.
- VRAM
- 24 GB
- Bandwidth
- 936 GB/s
- Fair used price
- $600–750
- Models that fit
- 124
| Device | VRAM | Bandwidth | 8B at Q4 | Fit | Price |
|---|---|---|---|---|---|
| Apple M3 Ultra | 512 GB unified | 819 GB/s | 43–90 tok/s | 158 | — |
| Apple M5 Ultra | 512 GB unified | 1,229 GB/s | 59–123 tok/s | 158 | — |
| Apple M2 Ultra | 192 GB unified | 800 GB/s | 43–89 tok/s | 147 | — |
| Apple M1 Ultra | 128 GB unified | 800 GB/s | 43–89 tok/s | 143 | — |
| Apple M3 Max | 128 GB unified | 400 GB/s | 24–49 tok/s | 143 | — |
| Apple M4 Max | 128 GB unified | 546 GB/s | 31–64 tok/s | 143 | — |
| Apple M5 Max | 128 GB unified | 614 GB/s | 34–71 tok/s | 143 | — |
| Apple M2 Max | 96 GB unified | 400 GB/s | 24–49 tok/s | 139 | — |
| Apple M3 Max (30-core GPU) | 96 GB unified | 300 GB/s | 18–38 tok/s | 139 | — |
| NVIDIA RTX PRO 6000 Blackwell | 96 GB | 1,792 GB/s | 113–234 tok/s | 143 | $8,565 |
| NVIDIA RTX PRO 6000 Blackwell (Max-Q Edition) | 96 GB | 1,792 GB/s | 113–234 tok/s | 143 | — |
| NVIDIA A100 80GB (PCIe) | 80 GB | 1,935 GB/s | 102–196 tok/s | 143 | $15,000 |
| NVIDIA H100 80GB (PCIe) | 80 GB | 2,000 GB/s | 120–250 tok/s | 143 | $25,000 |
| NVIDIA H100 80GB (SXM) | 80 GB | 3,350 GB/s | 157–327 tok/s | 143 | — |
| NVIDIA RTX PRO 5000 Blackwell (72 GB) | 72 GB | 1,344 GB/s | 94–194 tok/s | 139 | — |
| Apple M1 Max | 64 GB unified | 400 GB/s | 24–49 tok/s | 131 | — |
| Apple M5 Pro | 64 GB unified | 307 GB/s | 19–39 tok/s | 131 | — |
| NVIDIA L40 | 48 GB | 864 GB/s | 77–147 tok/s | 131 | — |
| NVIDIA L40S | 48 GB | 864 GB/s | 77–147 tok/s | 131 | $7,499 |
| AMD Radeon PRO W7900 | 48 GB | 864 GB/s | 50–103 tok/s | 131 | — |
| NVIDIA RTX 6000 Ada Generation | 48 GB | 960 GB/s | 83–159 tok/s | 131 | $6,799 |
| NVIDIA RTX A6000 | 48 GB | 768 GB/s | 51–98 tok/s | 131 | $4,649 |
| NVIDIA RTX PRO 5000 Blackwell (48 GB) | 48 GB | 1,344 GB/s | 94–194 tok/s | 131 | — |
| Apple M3 Pro | 36 GB unified | 153 GB/s | 9.7–20 tok/s | 124 | — |
| Apple M4 Max (32-core GPU) | 36 GB unified | 410 GB/s | 24–50 tok/s | 124 | — |
| Apple M5 Max (32-core GPU) | 36 GB unified | 460 GB/s | 27–56 tok/s | 124 | — |
| AMD Instinct MI50 32GB | 32 GB | 1,024 GB/s | — | 124 | — |
| Apple M1 Pro | 32 GB unified | 200 GB/s | 13–26 tok/s | 124 | — |
| Apple M2 Pro | 32 GB unified | 200 GB/s | 13–26 tok/s | 124 | — |
| Apple M4 | 32 GB unified | 120 GB/s | 7.7–16 tok/s | 124 | — |
| Apple M5 | 32 GB unified | 153 GB/s | 9.7–20 tok/s | 124 | — |
| Apple M6 | 32 GB unified | 170 GB/s | 11–22 tok/s | 124 | — |
| Intel Arc Pro B65 | 32 GB | 608 GB/s | 35–72 tok/s | 124 | — |
| Intel Arc Pro B70 | 32 GB | 608 GB/s | 35–72 tok/s | 124 | $949 |
| AMD Radeon AI PRO R9700 | 32 GB | 640 GB/s | 42–87 tok/s | 124 | $1,299 |
| NVIDIA RTX 5000 Ada Generation | 32 GB | 576 GB/s | 56–107 tok/s | 124 | — |
| NVIDIA GeForce RTX 5090 | 32 GB | 1,792 GB/s | 113–234 tok/s | 124 | $4,700 |
| NVIDIA RTX PRO 4500 Blackwell (Workstation Edition) | 32 GB | 896 GB/s | 70–145 tok/s | 124 | — |
| Apple M2 | 24 GB unified | 100 GB/s | 6.4–13 tok/s | 95 | — |
| Apple M3 | 24 GB unified | 100 GB/s | 6.4–13 tok/s | 95 | — |
| Apple M4 Pro | 24 GB unified | 273 GB/s | 17–35 tok/s | 95 | — |
| Intel Arc Pro B60 | 24 GB | 456 GB/s | 28–57 tok/s | 124 | — |
| NVIDIA A10 | 24 GB | 600 GB/s | 41–79 tok/s | 124 | — |
| NVIDIA L4 | 24 GB | 300 GB/s | 32–61 tok/s | 124 | — |
| NVIDIA GeForce RTX 3090 | 24 GB | 936 GB/s | 60–115 tok/s | 124 | $1,499 |
| NVIDIA GeForce RTX 3090 Ti | 24 GB | 1,008 GB/s | 63–122 tok/s | 124 | $1,999 |
| NVIDIA GeForce RTX 4090 | 24 GB | 1,008 GB/s | 86–165 tok/s | 124 | $1,599 |
| NVIDIA RTX A5000 | 24 GB | 768 GB/s | 51–98 tok/s | 124 | — |
| NVIDIA RTX PRO 4000 Blackwell | 24 GB | 672 GB/s | 56–116 tok/s | 124 | — |
| AMD Radeon RX 7900 XTX | 24 GB | 960 GB/s | 54–112 tok/s | 124 | $999 |
| NVIDIA Tesla P40 | 24 GB | 346 GB/s | — | 124 | — |
| AMD Radeon RX 7900 XT | 20 GB | 800 GB/s | 47–98 tok/s | 106 | $899 |
| Apple M1 | 16 GB unified | 68 GB/s | 4.4–9.2 tok/s | 77 | — |
| Intel Arc Pro B50 | 16 GB | 224 GB/s | 15–31 tok/s | 87 | $349 |
| NVIDIA GeForce RTX 4060 Ti 16GB | 16 GB | 288 GB/s | 31–59 tok/s | 87 | $499 |
| NVIDIA GeForce RTX 4070 Ti Super | 16 GB | 672 GB/s | 63–121 tok/s | 87 | $799 |
| NVIDIA GeForce RTX 4080 | 16 GB | 717 GB/s | 66–127 tok/s | 87 | $1,199 |
| NVIDIA GeForce RTX 4080 Super | 16 GB | 736 GB/s | 68–130 tok/s | 87 | $999 |
| NVIDIA GeForce RTX 5060 Ti 16GB | 16 GB | 448 GB/s | 40–83 tok/s | 87 | $805 |
| NVIDIA GeForce RTX 5070 Ti | 16 GB | 896 GB/s | 70–145 tok/s | 87 | $749 |
| NVIDIA GeForce RTX 5080 | 16 GB | 960 GB/s | 74–153 tok/s | 87 | $999 |
| NVIDIA RTX PRO 2000 Blackwell | 16 GB | 288 GB/s | 27–56 tok/s | 87 | — |
| AMD Radeon RX 6800 XT | 16 GB | 512 GB/s | — | 87 | — |
| AMD Radeon RX 6900 XT | 16 GB | 512 GB/s | — | 87 | — |
| AMD Radeon RX 7600 XT | 16 GB | 288 GB/s | 20–42 tok/s | 87 | — |
| AMD Radeon RX 7800 XT | 16 GB | 624 GB/s | 39–81 tok/s | 87 | $499 |
| AMD Radeon RX 7900 GRE | 16 GB | 576 GB/s | 36–76 tok/s | 87 | — |
| AMD Radeon RX 9060 XT 16GB | 16 GB | 320 GB/s | 24–49 tok/s | 87 | $349 |
| AMD Radeon RX 9070 | 16 GB | 640 GB/s | 42–87 tok/s | 87 | $549 |
| AMD Radeon RX 9070 XT | 16 GB | 640 GB/s | 42–87 tok/s | 87 | $599 |
| Intel Arc B580 | 12 GB | 456 GB/s | 28–57 tok/s | 77 | $249 |
| NVIDIA GeForce RTX 3060 (12GB) | 12 GB | 360 GB/s | 26–50 tok/s | 77 | $329 |
| NVIDIA GeForce RTX 3080 12GB | 12 GB | 912 GB/s | 59–112 tok/s | 77 | — |
| NVIDIA GeForce RTX 3080 Ti | 12 GB | 912 GB/s | 59–112 tok/s | 77 | $1,199 |
| NVIDIA GeForce RTX 4070 | 12 GB | 504 GB/s | 50–96 tok/s | 77 | $599 |
| NVIDIA GeForce RTX 4070 Super | 12 GB | 504 GB/s | 50–96 tok/s | 77 | $599 |
| NVIDIA GeForce RTX 4070 Ti | 12 GB | 504 GB/s | 50–96 tok/s | 77 | $799 |
| NVIDIA GeForce RTX 5070 | 12 GB | 672 GB/s | 56–116 tok/s | 77 | $549 |
| AMD Radeon RX 6700 XT | 12 GB | 384 GB/s | — | 77 | — |
| AMD Radeon RX 7700 XT | 12 GB | 432 GB/s | 29–60 tok/s | 77 | $449 |
| AMD Radeon RX 9070 GRE | 12 GB | 432 GB/s | 31–64 tok/s | 77 | — |
| NVIDIA GeForce GTX 1080 Ti | 11 GB | 484 GB/s | — | 77 | $699 |
| NVIDIA GeForce RTX 2080 Ti | 11 GB | 616 GB/s | — | 77 | $999 |
| Intel Arc B570 | 10 GB | 380 GB/s | 24–49 tok/s | 75 | $219 |
| NVIDIA GeForce RTX 3080 (10GB) | 10 GB | 760 GB/s | 50–97 tok/s | 75 | $699 |
| NVIDIA GeForce RTX 3060 Ti | 8 GB | 448 GB/s | 32–61 tok/s | 65 | $399 |
| NVIDIA GeForce RTX 3070 | 8 GB | 448 GB/s | 32–61 tok/s | 65 | $499 |
| NVIDIA GeForce RTX 3070 Ti | 8 GB | 608 GB/s | 42–80 tok/s | 65 | $599 |
| NVIDIA GeForce RTX 4060 | 8 GB | 272 GB/s | 29–56 tok/s | 65 | $299 |
| NVIDIA GeForce RTX 4060 Ti 8GB | 8 GB | 288 GB/s | 31–59 tok/s | 65 | $399 |
| NVIDIA GeForce RTX 5050 | 8 GB | 320 GB/s | 30–61 tok/s | 65 | $249 |
| NVIDIA GeForce RTX 5060 | 8 GB | 448 GB/s | 40–83 tok/s | 65 | $299 |
| NVIDIA GeForce RTX 5060 Ti 8GB | 8 GB | 448 GB/s | 40–83 tok/s | 65 | $379 |
| AMD Radeon RX 7600 | 8 GB | 288 GB/s | 20–42 tok/s | 65 | — |
| AMD Radeon RX 9060 XT 8GB | 8 GB | 320 GB/s | 24–49 tok/s | 65 | $299 |
Quick buying guide
- Under $300
- RTX 3060 12GB12 GB VRAM, runs all 7–8B models
- Under $500
- RTX 4060 Ti 16GBbest budget 16 GB card
- $500–700
- RTX 4070 Super or AMD RX 9070 XT
- $1,000–1,600
- RTX 4090 24GBfastest single consumer card; 32B-class models at Q4_K_M
- Mac laptop
- M3 Pro 36GB or M4 Prosilent, power-efficient
- Mac desktop
- M4 Max 128GB~96 GB usable: 70B-class models at Q4_K_M
Speed is Llama 3.1 8B at Q4_K_M: a bandwidth-roofline estimate shown as a range unless the device page marks it measured. Mac memory is the full unified pool; about 75% of it is usable by a model. Fit counts catalogued model variants whose Q4_K_M weights plus runtime overhead fit the usable memory, before any context is added. Methodology
Keep going
Tools