GPUs & Macs › NVIDIA A40
NVIDIA A40: local LLM speed
Llama 2 7B (Q4_0) generates 126 tokens/s on the NVIDIA A40 (CUDA), and processes prompts at 5,256 tokens/s.
Measured results
| Model | Generate tok/s | Prompt tok/s | Quant | Backend | Test | Source |
|---|---|---|---|---|---|---|
| Llama 2 7B | 126 | 5,256 | Q4_0 | CUDA FA on | tg128 | llama.cpp #15013 details |
| Llama 2 7B | 124 | 4,609 | Q4_0 | CUDA FA off | tg128 | llama.cpp #15013 details |
| Llama 3.2 1B Instruct | 284 | 12,156 | Q4_K_M | llamafile | LocalScore avg | LocalScore details |
| Llama 3.1 8B Instruct | 80.6 | 3,270 | Q4_K_M | llamafile | LocalScore avg | LocalScore details |
| Qwen2.5 14B Instruct | 46.2 | 1,825 | Q4_K_M | llamafile | LocalScore avg | LocalScore details |
What fits on the NVIDIA A40
Estimates for this chip are scaled from a measured run: Llama 2 7B Q4_0 at 126 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 | 16.6 estimate |
| gemma-4-12B-it | Q8_0 | 12.7 GB | 38.1 estimate |
| Qwen3.5-9B | Q8_0 | 9.5 GB | 50.7 estimate |
| Qwen3.5-4B | Q8_0 | 4.5 GB | 108 estimate |
| Qwen3.6-35B-A3B | Q8_0 | 36.9 GB | 76.3 estimate |
| gemma-4-26B-A4B-it | Q8_0 | 26.9 GB | 58.4 estimate |
| MiMo-V2.6-Distill-Qwen-9B | Q8_0 | 9.5 GB | 50.6 estimate |
| gemma-4-E4B-it | Q8_0 | 8.2 GB | 58.9 estimate |
| Muse-Glimmer-30B | Q8_0 | 29.6 GB | 16.3 estimate |
| Qwen3-4B | Q8_0 | 4.3 GB | 113 estimate |
| gemma-4-31B-it | Q8_0 | 32.6 GB | 14.8 estimate |
| gemma-4-26B-A4B-it-qat-q4_0-unquantized | UD-Q4_K_XL | 14.2 GB | 110 estimate |
| gemma-4-12B-it-qat-q4_0-unquantized | UD-Q4_K_XL | 6.7 GB | 71.9 estimate |
| gemma-3-1b-it | Q8_0 | 1.1 GB | 402 estimate |
Similar hardware
NVIDIA L40 · NVIDIA L40S · NVIDIA Tesla V100 32GB · NVIDIA A30 · NVIDIA Tesla P40 · NVIDIA Tesla M40 · NVIDIA L4 · NVIDIA A10