DeepSeek VRAM & GPU Requirement Calculator
Calculate VRAM requirements and GPU count for DeepSeek deployment. Support for NVIDIA, AMD, Apple, and Huawei
GPUs
Memory Requirements 884.98 GB
Requires 12 GPUs (based on memory capacity)
880 GB
All model weights
0.04 GB
Conversation history cache
4.12 GB
Expert model optimization
0.82 GB
Temporary computation cache
Performance Metrics
Visualize 11990 tok/s →Total Throughput
11990 tokens/s
Throughput Per User
11990 tokens/s/user
Estimated Latency
13 ms
Max QPS
79.9 queries/s
Estimated using memory bandwidth and tensor compute throughput across 12x NVIDIA H100 (80GB).
Quick Deploy CLI Command
vllm serve deepseek-ai/DeepSeek-V4-Pro --tensor-parallel-size 12 --max-model-len 8192
Flags reflect your selected precision, context length, and tensor parallel GPU count.Open DeepSeek-V4-Pro spec page →
DeepSeek Architecture & Memory Profile
Total Parameters
1600B
Architecture
MoE (49B active)
Developer / Provider
DeepSeek
Recommended Precision
FP4+FP8
DeepSeek VRAM Matrix by Precision & Context Length
Precomputed single-user memory requirements (model weights + KV cache + runtime overhead).
| Precision | 8K Context | 32K Context | 128K Context |
|---|---|---|---|
| FP16 | 3205.0 GB | 3207.2 GB | 3210.2 GB |
| FP8 | 1605.0 GB | 1607.1 GB | 1609.5 GB |
| INT8 | 1605.0 GB | 1607.1 GB | 1609.5 GB |
| INT4 | 805.0 GB | 807.0 GB | 809.2 GB |
Recommended GPU Hardware for DeepSeek (FP4+FP8)
Minimum GPU count for 8K context at batch size 1 and estimated generation throughput.
| GPU Hardware | Min GPUs | Total VRAM Pool | Est. Speed |
|---|---|---|---|
| NVIDIA RTX 4090 (24GB) | 37× | 888 GB | 31555 tok/s |
| NVIDIA RTX 5090 (32GB) | 28× | 896 GB | 33196 tok/s |
| Apple M4 Max (128GB) (128GB) | 7× | 896 GB | 539 tok/s |
| NVIDIA A100 (80GB) (80GB) | 12× | 960 GB | 7049 tok/s |
| NVIDIA H100 (80GB) | 12× | 960 GB | 11990 tok/s |
| NVIDIA H200 (141GB) | 7× | 987 GB | 6305 tok/s |
| Huawei Ascend 910B (64GB) | 14× | 896 GB | 5935 tok/s |
DeepSeek Deployment FAQ
How much VRAM is needed to run DeepSeek?
At 8K context and batch size 1, DeepSeek uses about 3205.0 GB in FP16, 1605.0 GB in FP8, and 805.0 GB in INT4/AWQ quantization.
Can I run DeepSeek on RTX 4090 or H100 GPUs?
Running DeepSeek in FP4+FP8 precision requires at least 37x RTX 4090 (24GB) or 12x NVIDIA H100 (80GB) GPUs.
What precision should I choose for DeepSeek in production?
For DeepSeek (1600B by DeepSeek), FP4+FP8 provides the best balance between inference speed, VRAM footprint, and output quality.