GPUs & Macs › NVIDIA P104-100
NVIDIA P104-100: local LLM speed
Llama 2 7B (Q4_0) generates 48.6 tokens/s on the NVIDIA P104-100 (Vulkan), and processes prompts at 325 tokens/s.
Measured results
| Model | Generate tok/s | Prompt tok/s | Quant | Backend | Test | Source |
|---|---|---|---|---|---|---|
| Llama 2 7B | 48.6 | 325 | Q4_0 | Vulkan FA on | tg128 | llama.cpp #10879 details |
| Llama 2 7B | 46.2 | 312 | Q4_0 | Vulkan FA off | tg128 | llama.cpp #10879 details |
What fits on the NVIDIA P104-100
Estimates for this chip are scaled from a measured run: Llama 2 7B Q4_0 at 48.6 tok/s (Vulkan, source). Fit assumes the whole model in GPU memory with ~8K context.
| Model | Best quant that fits | File | Expected tok/s |
|---|---|---|---|
| Qwen3.5-4B | IQ4_XS | 2.5 GB | 75.0 estimate |
| Qwen3-4B | Q4_K_M | 2.5 GB | 74.4 estimate |
| gemma-3-1b-it | Q8_0 | 1.1 GB | 155 estimate |
| DeepSeek-R1-0528-Qwen3-8B | UD-IQ1_M | 2.4 GB | 77.8 estimate |
| Llama-3.2-3B-Instruct | Q5_K_M | 2.3 GB | 80.1 estimate |
| Phi-4-mini-instruct | Q4_K_M | 2.5 GB | 74.6 estimate |
| gemma-3-4b-it | Q4_K_M | 2.5 GB | 74.7 estimate |
| Qwen3.5-0.8B | Q8_0 | 0.8 GB | 155 estimate |
| Llama-3.2-3B-Instruct-uncensored | Q5_K_M | 2.6 GB | 71.7 estimate |
| Qwen3-0.6B | Q8_0 | 0.8 GB | 155 estimate |
| GLM-OCR | Q8_0 | 1.0 GB | 155 estimate |
| Qwen3-ASR-1.7B | Q8_0 | 2.2 GB | 85.9 estimate |
| SmolLM3-3B | Q4_K_M | 1.9 GB | 97.1 estimate |
| Nanbeige4.2-3B | Q4_K_S | 2.6 GB | 72.8 estimate |
Similar hardware
NVIDIA P102-100 · NVIDIA P106-100 · NVIDIA Tesla P4 · NVIDIA Tesla K80 · NVIDIA Tesla P100 · NVIDIA Tesla T4 · NVIDIA Tesla V100 16GB · NVIDIA A30