H100 80GB for local LLMs
5 measured runs. 80 GB of VRAM; memory measured at 3252 GB/s, which sets the roofline for every model below.
| Model | Quant | Engine | tok/s | Roofline | Share |
|---|---|---|---|---|---|
| Qwen3-14B | AWQ | vLLM | 147.6 | 380.3 | 39% |
| Qwen3-14B | Q4_K_M | llama.cpp | 144.4 | 374.3 | 39% |
| Qwen3-8B | Q4_K_M | llama.cpp | 227.0 | 684.4 | 33% |
| Qwen3.8-27B | W4A16-AWQ | vLLM | 87.8 | 211.9 | 41% |
| Qwen3.8-27B | UD-Q4_K_M | llama.cpp | 83.6 | 205.7 | 41% |
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.