L40S for local LLMs
4 measured runs. 45 GB of VRAM; memory measured at 762 GB/s, which sets the roofline for every model below.
| Model | Quant | Engine | tok/s | Roofline | Share |
|---|---|---|---|---|---|
| Qwen3-14B | AWQ | vLLM | 77.7 | 89.4 | 87% |
| Qwen3-14B | Q4_K_M | llama.cpp | 74.1 | 88.1 | 84% |
| Qwen3.8-27B | W4A16-AWQ | vLLM | 43.6 | 48.9 | 89% |
| Qwen3.8-27B | UD-Q4_K_M | llama.cpp | 40.0 | 48.4 | 83% |
Share = measured tok/s over the roofline: the bytes a token reads over this GPU's measured memory speed. Open a quant for the test machine and the per-role and per-kernel breakdown.