A40

Speed is modeled, not measured: decode is memory-bandwidth bound, so tokens per second is bytes read per token against achievable bandwidth. Mixture-of-experts models carry a wider band because only the routed experts are read each step, and few have been measured publicly.

Aggregated from community-submitted runs, so the spread is wide by nature — it covers different models, resolutions, step counts and settings, not one controlled configuration. Read the middle 50% rather than the median alone. These figures are reproduced with attribution from llama.cpp-discussion-15013.

What AI models can a A40 run?

1857 of 2118 indexed open-weight models fit a A40 at 131,072 context with f16 KV cache, the largest being Qwen2.5-Coder-14B-Instruct at Q5_K_M. That covers text, vision-language, image, video and speech models.

How much usable memory does a A40 actually have?

Its nameplate is 48 GB, but about 44.64 GiB is available to a model once driver and compositor overhead is accounted for.

Is a A40 fast for local AI?

Its memory bandwidth is 696 GB/s, and that figure — not teraflops — is what governs token generation speed. Capacity decides what you can run; bandwidth decides how fast it runs.