Datasets / Local AI GPU Value Index
Local AI GPU Value Index
The open dataset behind our value rankings: every discrete GPU we track with its launch MSRP, VRAM, memory bandwidth, and — where independent test data exists — measured llama.cpp generation tokens/s per $1000. Every number carries a source link. License: CC BY 4.0.
Download
Download CSV Download JSON Version v2026-09-04 · price basis: launch MSRP · license CC BY 4.0
Cite the versioned URL for stable references: https://compareaihardware.com/datasets/gpu-value-index/gpu-value-index-v2026-09-04.csv. See full methodology.
Headline: measured tokens/s per $1000
Ranked by measured generation throughput per $1000 of launch MSRP. Cohort: Qwen3-8B @ Q4_K_XL (19 GPUs measured). Switch cohort: Llama-3.1-8B @ Q4_K_M
| # | GPU | tok/s | tok/s per $1000 | VRAM | MSRP |
|---|---|---|---|---|---|
| 1 | Arc B580 | 41.0 | 164.66 | 12 GB | $249 |
| 2 | GeForce RTX 4070 SUPER | 92.0 | 153.59 | 12 GB | $599 |
| 3 | GeForce RTX 5080 | 132.0 | 132.13 | 16 GB | $999 |
| 4 | GeForce RTX 3060 12GB | 38.0 | 115.50 | 12 GB | $329 |
| 5 | GeForce RTX 5090 | 215.0 | 107.55 | 32 GB | $1,999 |
| 6 | GeForce RTX 4080 SUPER | 102.0 | 102.10 | 16 GB | $999 |
| 7 | GeForce RTX 4090 | 125.0 | 78.17 | 24 GB | $1,599 |
| 8 | Radeon RX 7900 XTX | 76.0 | 76.08 | 24 GB | $999 |
| 9 | GeForce RTX 3090 | 85.0 | 56.70 | 24 GB | $1,499 |
| 10 | RTX 6000 Ada | 110.0 | 16.18 | 48 GB | $6,800 |
Cohort shown: Llama-3.1-8B @ Q4_K_M. The two cohorts use different models AND different quantizations — they are never averaged or normalized against each other.
Spec tier: bandwidth & VRAM per $1000 (all 57 GPUs)
A coverage metric computed from vendor specs ÷ launch MSRP — a proxy, not performance. Full column set lives in the download; top entries by bandwidth per $1000:
| GPU | GB/s per $1000 | GB per $1000 | Bandwidth | VRAM | MSRP |
|---|---|---|---|---|---|
| Arc A580 | 2,860.34 | 44.69 | 512 GB/s | 8 GB | $179 |
| Arc B580 | 1,831.33 | 48.19 | 456 GB/s | 12 GB | $249 |
| Arc A750 | 1,771.63 | 27.68 | 512 GB/s | 8 GB | $289 |
| Arc B570 | 1,735.16 | 45.66 | 380 GB/s | 10 GB | $219 |
| Arc A770 16GB | 1,604.58 | 45.85 | 560 GB/s | 16 GB | $349 |
| GeForce RTX 5060 | 1,498.33 | 26.76 | 448 GB/s | 8 GB | $299 |
| GeForce RTX 5050 | 1,285.14 | 32.13 | 320 GB/s | 8 GB | $249 |
| Radeon RX 7800 XT | 1,250.50 | 32.06 | 624 GB/s | 16 GB | $499 |
| Radeon RX 7900 XT | 1,232.67 | 30.82 | 800 GB/s | 20 GB | $649 |
| GeForce RTX 5070 | 1,224.04 | 21.86 | 672 GB/s | 12 GB | $549 |
| GeForce RTX 5070 Ti | 1,196.26 | 21.36 | 896 GB/s | 16 GB | $749 |
| Radeon RX 9070 | 1,165.76 | 29.14 | 640 GB/s | 16 GB | $549 |
| GeForce RTX 3060 12GB | 1,094.22 | 36.47 | 360 GB/s | 12 GB | $329 |
| GeForce RTX 3080 10GB | 1,087.27 | 14.31 | 760 GB/s | 10 GB | $699 |
| Radeon RX 7600 | 1,070.63 | 29.74 | 288 GB/s | 8 GB | $269 |
Evidence legend
- vendor_spec — manufacturer-published hardware specification (VRAM GB, memory bandwidth GB/s).
- manufacturer_pricing — launch MSRP in USD at product introduction. Never street or used-market prices.
- independent_benchmark — third-party measured llama.cpp generation throughput, non-estimate rows only, each with source URL and test date.
Coverage
57 GPUs total · 42 with sourced launch MSRP · measured cohorts: Qwen3-8B @ Q4_K_XL = 20 GPUs, Llama-3.1-8B @ Q4_K_M = 10 GPUs.
How to cite
CompareAIHardware, "Local AI GPU Value Index", dataset version v2026-09-04, September 2026. Licensed under CC BY 4.0. Available at: https://compareaihardware.com/datasets/gpu-value-index