GPUs & Macs › NVIDIA H100 SXM
NVIDIA H100 SXM: local LLM speed
Llama 2 7B (Q4_0) generates 281 tokens/s on the NVIDIA H100 SXM (CUDA), and processes prompts at 11,263 tokens/s.
Measured results
| Model | Generate tok/s | Prompt tok/s | Quant | Backend | Test | Source |
|---|---|---|---|---|---|---|
| Llama 2 7B | 281 | 11,263 | Q4_0 | CUDA FA on | tg128 | llama.cpp #15013 details |
| Llama 2 7B | 268 | 9,918 | Q4_0 | CUDA FA off | tg128 | llama.cpp #15013 details |
| Llama 3.2 1B Instruct | 96.2 | 14,608 | Q4_K_M | llamafile | LocalScore avg | LocalScore details |
| Llama 3.1 8B Instruct | 42.4 | 4,687 | Q4_K_M | llamafile | LocalScore avg | LocalScore details |
| Qwen2.5 14B Instruct | 23.4 | 2,609 | Q4_K_M | llamafile | LocalScore avg | LocalScore details |
What fits on the NVIDIA H100 SXM
Estimates for this chip are scaled from a measured run: Llama 2 7B Q4_0 at 281 tok/s (CUDA, source). Fit assumes the whole model in GPU memory with ~8K context.
| Model | Best quant that fits | File | Expected tok/s |
|---|---|---|---|
| Qwen3.8-27B | Q8_0 | 29.0 GB | 36.9 estimate |
| gemma-4-12B-it | Q8_0 | 12.7 GB | 84.7 estimate |
| Qwen3.5-9B | Q8_0 | 9.5 GB | 113 estimate |
| Qwen3.5-4B | Q8_0 | 4.5 GB | 239 estimate |
| Qwen3.6-35B-A3B | Q8_0 | 36.9 GB | 170 estimate |
| Qwen3.8-Flash-Next | UD-IQ1_M | 74.5 GB | 14.4 estimate |
| gemma-4-26B-A4B-it | Q8_0 | 26.9 GB | 130 estimate |
| MiMo-V2.6-Distill-Qwen-9B | Q8_0 | 9.5 GB | 112 estimate |
| gemma-4-E4B-it | Q8_0 | 8.2 GB | 131 estimate |
| Muse-Glimmer-30B | Q8_0 | 29.6 GB | 36.2 estimate |
| Qwen3-4B | Q8_0 | 4.3 GB | 251 estimate |
| gemma-4-31B-it | Q8_0 | 32.6 GB | 32.9 estimate |
| gemma-4-26B-A4B-it-qat-q4_0-unquantized | UD-Q4_K_XL | 14.2 GB | 245 estimate |
| gemma-4-12B-it-qat-q4_0-unquantized | UD-Q4_K_XL | 6.7 GB | 160 estimate |
Similar hardware
NVIDIA A100 80GB PCIe · NVIDIA H100 PCIe · NVIDIA A100 80GB SXM · NVIDIA L40 · NVIDIA A40 · NVIDIA L40S · NVIDIA Tesla V100 32GB · NVIDIA A30