Home › GPUs & Macs › Value per dollar
GPU value for local LLMs: tokens per second per dollar
Measured generation speed on Llama 2 7B Q4_0 (the standard llama.cpp test, best backend for each card) divided by the card's launch price. 50 cards have both numbers.
These are launch MSRPs, not used prices. Used prices are pending: we have not found a used-price source we can republish with a date and a link. Used and street prices can sit well below or well above launch price, so treat this table as a starting point and check current listings.
As an Amazon Associate I earn from qualifying purchases. Amazon links are searches for the card name and carry our Associates tag; they never change the order, which is sorted by measured speed per launch dollar. Affiliate policy.
How the numbers are worked out
tok/s per $100 = measured tok/s / launch MSRP × 100 GB per $100 = VRAM in GB / launch MSRP × 100
Speed is the fastest measured run of Llama 2 7B Q4_0 for the card with the whole model in GPU memory; click it for every run and its source. The test model is small (a 3.8 GB file), so this compares memory bandwidth and software support more than capacity. For bigger models, VRAM decides what runs at all: use the GB per $100 column and the Will it run? calculator. Thinking of renting instead? See rent or buy.
Prices: launch MSRP in US dollars from the Wikipedia series articles at these revisions, checked 2026-10-22. Those tables cite the vendors' announcements. Where a card launched at two prices (standard and Founders Edition) we use the standard price and say so next to it.
- GeForce RTX 50 series, revision 1380910384
- GeForce RTX 40 series, revision 1374808024
- GeForce RTX 30 series, revision 1374808230
- GeForce RTX 20 series, revision 1378260311
- GeForce 10 series, revision 1379452519
- GeForce 16 series, revision 1287210349
- Radeon RX 9000 series, revision 1379257521
- Radeon RX 7000 series, revision 1356272510
- Radeon RX 6000 series, revision 1347872354
- Intel Arc, revision 1379272005
Priced but no measured Llama 2 7B Q4_0 run yet: Radeon RX 6950 XT, Radeon RX 7700 XT, GTX 1080, RTX 2060, RTX 2070, RTX 2080, RTX 2080 SUPER, RTX 3050, RTX 3070 Ti, RTX 3080 Ti, RTX 4060, RTX 4060 Ti 16GB, RTX 5060, TITAN RTX. Submit one.