GPUFits
GPU Hub: Local LLM Hardware Overview
Each GPU page includes specs, commentary from an LLM-inference perspective, value metrics (bandwidth per dollar, VRAM per dollar), and a verdict matrix with theoretical speed for 14 mainstream models at Q4_K_M @8K. Used price is preferred; MSRP is shown when no used price exists.
| GPU | VRAM | Bandwidth | Reference price | Bandwidth per $ |
|---|---|---|---|---|
| H100 80GB | 80 GB | 3350 GB/s | $30,000 | 0.11 |
| A100 80GB | 80 GB | 2039 GB/s | $15,000 | 0.14 |
| RTX 5090 | 32 GB | 1792 GB/s | $1,999 | 0.90 |
| RTX 4090 | 24 GB | 1008 GB/s | $2,400vs last month +118% | 0.42 |
| RX 7900 XTX | 24 GB | 960 GB/s | $650 | 1.48 |
| RTX 3090 | 24 GB | 936 GB/s | $1,100vs last month +10% | 0.85 |
| Mac Studio M3 Ultra (96GB) | 96 GB* | 819 GB/s | $5,299 | 0.15 |
| RTX A6000 | 48 GB | 768 GB/s | $2,800 | 0.27 |
| RTX 4070 Ti Super | 16 GB | 672 GB/s | $760vs last month +38% | 0.88 |
| Mac Studio M4 Max (64GB) | 64 GB* | 546 GB/s | $2,899 | 0.19 |
| RTX 3060 12GB | 12 GB | 360 GB/s | $260vs last month +44% | 1.38 |
| Mac mini M4 Pro (48GB) | 48 GB* | 273 GB/s | $1,799 | 0.15 |
Apple unified memory is counted at 75% usable for models (the rest serves the system/display).
Data verified 2026-09-01