H200 NVL for local LLMs
2 measured runs. 140 GB of VRAM; memory measured at 4513 GB/s, which sets the roofline for every model below.
| Model | Quant | Engine | tok/s | Roofline | Share |
|---|---|---|---|---|---|
| Qwen3-14B | Q4_K_M | llama.cpp | 135.0 | 521.8 | 26% |
| Qwen3.8-27B | UD-Q4_K_M | llama.cpp | 78.6 | 286.7 | 27% |
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.