GPU Value Index for Local LLMs
Estimated tokens/sec per dollar, every tracked card ranked. Open methodology, CC BY 4.0 data, dated prices.
Last updated: August 16, 2026 · Editor: ModelFit Team
- Best value overall: AMD Radeon RX 7900 XT at ~151 tok/s per $1,000 (used market).
- Best value new NVIDIA: NVIDIA GeForce RTX 5070 Ti at ~84 tok/s per $1,000.
- Fastest regardless of price: NVIDIA GeForce RTX 5090 at ~162 tok/s est. (8B Q4).
Cite this page: ModelFit, GPU Value Index for Local LLMs, https://modelfit.io/gpu/value-index/, updated 2026-08-16, CC BY 4.0. Speed figures are engine estimates, not measurements.
The ranking
Method: tokens/sec are ModelFit engine estimates for an 8B Q4_K_M model at 16k context (bandwidth-bound, not measured). Value = est. tok/s / dated median listing price x 1,000. AMD scores assume a working ROCm setup. The unified-memory row is a complete system, not a card. Prices: see each card page for the live listing link.
FAQ
What is the best GPU value for local LLMs in 2026?
On estimated speed per dollar, the AMD Radeon RX 7900 XT leads this index at ~151 tokens/sec per $1,000 (83 tok/s est. on an 8B Q4 model at ~$550 used). That is a used-market price; among new NVIDIA cards, the NVIDIA GeForce RTX 5070 Ti leads at ~84 tok/s per $1,000.
Is the RTX 3090 still worth buying for local LLMs?
On value, yes: 24 GB of VRAM at ~$900 used gives it one of the best tokens-per-dollar scores among NVIDIA cards (~108 tok/s per $1,000). Newer 16 GB cards are faster per watt but hold far less of a model in VRAM.
Are AMD cards good for running local LLMs?
On paper they top the value table, but the estimate assumes a working ROCm setup (Linux, or Windows via recent drivers). If you want zero-friction Ollama, NVIDIA remains the safer buy; if you enjoy tuning, the RX 7900 series offers the most VRAM per dollar.
How is this index computed?
Tokens/sec figures are ModelFit engine estimates for an 8B Q4_K_M model at 16k context (bandwidth-bound model, not measurements). Value = estimated tokens/sec divided by the dated median listing price, scaled per $1,000. Prices are medians of legitimate listings, refreshed periodically; used-market cards are labeled. Data: CC BY 4.0.