GPUs & Macs › NVIDIA Tesla P4
NVIDIA Tesla P4: local LLM speed
Llama 2 7B (Q4_0) generates 33.3 tokens/s on the NVIDIA Tesla P4 (CUDA), and processes prompts at 515 tokens/s.
Measured results
| Model | Generate tok/s | Prompt tok/s | Quant | Backend | Test | Source |
|---|---|---|---|---|---|---|
| Llama 2 7B | 33.3 | 515 | Q4_0 | CUDA FA off | tg128 | llama.cpp #15013 details |
| Llama 2 7B | 33.1 | 530 | Q4_0 | CUDA FA on | tg128 | llama.cpp #15013 details |
| Llama 2 7B | 28.4 | 287 | Q4_0 | Vulkan FA on | tg128 | llama.cpp #10879 details |
| Llama 2 7B | 28.0 | 266 | Q4_0 | Vulkan FA off | tg128 | llama.cpp #10879 details |
What fits on the NVIDIA Tesla P4
Estimates for this chip are scaled from a measured run: Llama 2 7B Q4_0 at 33.3 tok/s (CUDA, source). Fit assumes the whole model in GPU memory with ~8K context.
| Model | Best quant that fits | File | Expected tok/s |
|---|---|---|---|
| gemma-4-12B-it | IQ4_XS | 6.4 GB | 20.0 estimate |
| Qwen3.5-9B | Q5_K_S | 6.4 GB | 20.0 estimate |
| Qwen3.5-4B | Q8_0 | 4.5 GB | 28.4 estimate |
| MiMo-V2.6-Distill-Qwen-9B | Q4_K_M | 5.8 GB | 21.8 estimate |
| gemma-4-E4B-it | Q5_K_M | 5.5 GB | 23.2 estimate |
| Qwen3-4B | Q8_0 | 4.3 GB | 29.7 estimate |
| gemma-3-1b-it | Q8_0 | 1.1 GB | 106 estimate |
| gemma-4-E2B-it | Q8_0 | 5.0 GB | 25.6 estimate |
| DeepSeek-R1-0528-Qwen3-8B | Q5_K_M | 5.9 GB | 21.8 estimate |
| Llama-3.2-3B-Instruct | Q8_0 | 3.4 GB | 37.2 estimate |
| Phi-4-mini-instruct | Q8_0 | 4.1 GB | 31.2 estimate |
| gemma-3-4b-it | Q8_0 | 4.1 GB | 30.8 estimate |
| Qwen3-14B | UD-IQ3_XXS | 6.0 GB | 21.2 estimate |
| gemma-3-12b-it | Q3_K_M | 6.0 GB | 21.2 estimate |
Similar hardware
NVIDIA P106-100 · NVIDIA P102-100 · NVIDIA Tesla K80 · NVIDIA P104-100 · NVIDIA Tesla P100 · NVIDIA Tesla T4 · NVIDIA Tesla V100 16GB · NVIDIA A30