The NVIDIA Tesla P4 is an enterprise-class GPU built on the Pascal architecture, featuring 8 GB of memory with 192 GB/s of memory bandwidth and a 75 W TDP. It delivers 5.5 TFLOPS of FP32 performance. Currently available from 2 providers starting at $0.04/GPU/hour, with a market median of $0.05/GPU/hour across 2 configurations.
Hardware specifications
Same across all providers
Median price across all providers
2 offerings from 2 providers
Sorted by price ascending. Compare configurations side-by-side.
LLMs that fit on the NVIDIA Tesla P4
Estimated VRAM for popular open-weight LLMs at 16-bit (FP16/BF16), 8-bit (FP8/INT8) and 4-bit (INT4/MXFP4) precision, against 8 GB per GPU.
| Model | Parameters | 16-bit | 8-bit | 4-bit |
|---|---|---|---|---|
| Gemma 3 12B Google | 12.2B | 29 GB Too large | 16 GB Too large | 9.4 GB Tight on 1× · 2% spare |
| Granite 4.1 8B IBM | 8.8B | 21 GB Too large | 12 GB Too large | 6.8 GB 1× · 1.2 GB free |
| Qwen3 8B Alibaba | 8.2B | 20 GB Too large | 11 GB Too large | 6.3 GB 1× · 1.7 GB free |
| Llama 3.1 8B Instruct Meta | 8B | 19 GB Too large | 11 GB Too large | 6.2 GB 1× · 1.8 GB free |
| Mistral 7B Instruct v0.3 Mistral AI | 7.2B | 17 GB Too large | 9.6 GB Tight on 1× · 0% spare | 5.6 GB 1× · 2.4 GB free |
| Qwen3 4B Alibaba | 4B | 9.7 GB Too large | 5.3 GB 1× · 2.7 GB free | 3.1 GB 1× · 4.9 GB free |
Estimates: model weights plus 20% for KV cache and runtime overhead, at short context lengths. Long contexts and large batches need more. Rows marked “tight” hold the weights but not that full margin. A ×N figure is the total VRAM across N of these GPUs and makes no claim about interconnect throughput. How we estimate this · Start from a model instead