Written by Jakub Rusinowski · Last updated July 21, 2026
Ranked for OCR and document scanning: Turning scanned pages, screenshots and photographed documents into accurate text.
Best overall: AMD Radeon RX 7900 XTX
24 GB VRAM at 960 GB/s. It runs 20 of the models that qualify for this workload; the strongest is Gemma 4 27B ⭐ at an estimated 38.3 tokens/sec.
| GPU | VRAM | MSRP | Models that fit | Best model it runs | Est. speed | |
|---|---|---|---|---|---|---|
| Best overall | AMD Radeon RX 7900 XTX | 24 GB | $999 | 20 | Gemma 4 27B ⭐ | ~38.3 tok/s |
| Best value | Intel Arc B570 | 10 GB | $219 | 11 | Qwen 3.5 14B | ~29.1 tok/s |
| Budget pick | Intel Arc B570 | 10 GB | $219 | 11 | Qwen 3.5 14B | ~29.1 tok/s |
| Most memory | NVIDIA DGX Spark | 128 GB | $4,699 | 25 | Qwen 2.5 VL 72B Instruct | ~4.4 tok/s |
| GPU | Score | VRAM | MSRP | Models fit | Est. speed | Tok/s per watt | Cost per model |
|---|---|---|---|---|---|---|---|
| AMD Radeon RX 7900 XTX | 79.3 | 24 GB | $999 | 20 | ~38.3 | 0.11 | $50 |
| NVIDIA GeForce RTX 4090 | 78.6 | 24 GB | $1,599 | 20 | ~40.1 | 0.09 | $80 |
| NVIDIA GeForce RTX 5090 | 77.6 | 32 GB | $1,999 | 20 | ~66.6 | 0.12 | $100 |
| NVIDIA GeForce RTX 3090 | 76.9 | 24 GB | $1,499 | 20 | ~37.4 | 0.11 | $75 |
| AMD Radeon RX 7900 XT | 74.8 | 20 GB | $899 | 16 | ~32.4 | 0.1 | $56 |
| NVIDIA GeForce RTX 5070 Ti | 73.7 | 16 GB | $749 | 13 | ~40.9 | 0.14 | $58 |
| NVIDIA GeForce RTX 5080 | 71.7 | 16 GB | $999 | 13 | ~43.5 | 0.12 | $77 |
| AMD Radeon RX 9070 XT | 70 | 16 GB | $599 | 13 | ~33.3 | 0.15 | $46 |
| Intel Arc B580 | 69.8 | 12 GB | $249 | 12 | ~34.5 | 0.18 | $21 |
| Intel Arc B570 | 68.1 | 10 GB | $219 | 11 | ~29.1 | 0.19 | $20 |
| AMD Radeon RX 9070 | 67.6 | 16 GB | $549 | 13 | ~30 | 0.14 | $42 |
| AMD Radeon RX 7800 XT | 67.4 | 16 GB | $499 | 13 | ~29.3 | 0.11 | $38 |
The AMD Radeon RX 7900 XTX — 24 GB of VRAM runs 20 qualifying models, the strongest being Gemma 4 (Legacy Listing — Unverified).
The Intel Arc B570 at $219, which runs 11 qualifying models.
8 GB is the entry point at which a model for this workload will run at all. More memory buys a stronger model, not just a faster one.