Table of Contents
Test Methodology
We benchmarked DeepSeek’s open-weight models across six GPUs on GigaGPU dedicated servers. Tests used vLLM with continuous batching. Each benchmark generates 512 output tokens from a 256-token prompt. DeepSeek models use a similar architecture to LLaMA but with mixture-of-experts (MoE) variants for larger models. For interactive results, visit our tokens/sec benchmark tool.
DeepSeek-R1 8B Throughput by GPU
DeepSeek-R1 8B is DeepSeek’s reasoning-focused model, competitive with LLaMA 3 8B on coding and math tasks. We benchmarked at FP16 precision with batch size 1 (single user).
| GPU | VRAM | FP16 tok/s | TTFT (ms) | Latency (512 tok) | Server $/hr |
|---|---|---|---|---|---|
| RTX 5090 | 32 GB | 132 | 35 | 3.9 sec | $1.80 |
| RTX 5080 | 16 GB | 81 | 52 | 6.3 sec | $0.85 |
| RTX 3090 | 24 GB | 59 | 58 | 8.7 sec | $0.45 |
| RTX 4060 Ti | 16 GB | 45 | 78 | 11.4 sec | $0.35 |
| RTX 4060 | 8 GB | 33 | 102 | 15.5 sec | $0.20 |
| RTX 3050 | 8 GB | 17 | 198 | 30.1 sec | $0.10 |
DeepSeek-R1 8B runs approximately 5% slower than LLaMA 3 8B on the same hardware due to its slightly different architecture. The difference is negligible in practice. For LLaMA-specific numbers, see our LLaMA 3 8B benchmark.
Concurrent User Throughput
DeepSeek-R1 8B under concurrent load with vLLM continuous batching:
| GPU | 1 User (tok/s) | 4 Users (total tok/s) | 8 Users (total tok/s) | 16 Users (total tok/s) |
|---|---|---|---|---|
| RTX 5090 | 132 | 398 | 645 | 875 |
| RTX 5080 | 81 | 242 | 390 | 495 |
| RTX 3090 | 59 | 175 | 270 | 340 |
| RTX 4060 Ti | 45 | 132 | 205 | 250 |
| RTX 4060 | 33 | 90 | 128 | 148 |
| RTX 3050 | 17 | 40 | 52 | 58 |
Scaling patterns mirror LLaMA 3 8B. The RTX 3090 handles 8 concurrent users comfortably with 270 total tok/s. For high-traffic deployments, consider an RTX 5090 or multi-GPU clusters.
Quantisation Impact
DeepSeek-R1 8B on the RTX 3090 at various precisions:
| Precision | VRAM Used | tok/s (bs=1) | tok/s (bs=8) | Quality Impact |
|---|---|---|---|---|
| FP16 | ~16 GB | 59 | 270 | Baseline |
| AWQ 4-bit | ~6 GB | 74 | 365 | Minimal |
| GPTQ 4-bit | ~6 GB | 71 | 350 | Minimal |
| FP8 | ~9 GB | 68 | 320 | Negligible |
AWQ 4-bit quantisation boosts throughput by ~25% while fitting the model on 8 GB GPUs. For quantisation format details, see our GPTQ vs AWQ vs GGUF guide.
Cost per Million Tokens
| GPU | FP16 Cost/1M Tokens | AWQ 4-bit Cost/1M Tokens | DeepSeek API Cost/1M Tokens |
|---|---|---|---|
| RTX 5090 | $3.79 | $2.95 | $2.00 (output) |
| RTX 5080 | $2.91 | $2.25 | $2.00 |
| RTX 3090 | $2.12 | $1.69 | $2.00 |
| RTX 4060 Ti | $2.16 | $1.58 | $2.00 |
| RTX 4060 | $1.68 | $1.14 | $2.00 |
| RTX 3050 | $1.63 | $1.05 | $2.00 |
Self-hosted DeepSeek beats API pricing on all GPUs with AWQ quantisation. The RTX 3090 at FP16 is roughly equivalent to API pricing but provides full data privacy and no rate limits. See our GPU vs API cost breakdown and cost calculator for detailed analysis.
DeepSeek vs LLaMA 3 vs Mistral: Speed Comparison
All three models at comparable parameter counts on the RTX 3090 (FP16, bs=1):
| Model | Parameters | tok/s (RTX 3090) | VRAM | Strengths |
|---|---|---|---|---|
| LLaMA 3 8B | 8B | 62 | ~16 GB | General purpose, broad ecosystem |
| Mistral 7B | 7B | 68 | ~14 GB | Fastest, sliding window attention |
| DeepSeek-R1 8B | 8B | 59 | ~16 GB | Strong reasoning, coding |
Mistral 7B is the fastest due to fewer parameters and optimised attention. LLaMA 3 8B and DeepSeek-R1 8B are closely matched. Model choice should be driven by task quality rather than speed differences. For Mistral benchmarks, see our Mistral 7B benchmark. For a complete overview, see best GPU for LLM inference.
GPU Recommendations for DeepSeek
Best overall: RTX 3090. At 59 tok/s (FP16) and $2.12 per million tokens, the RTX 3090 matches DeepSeek API pricing while keeping data private and eliminating rate limits. The 24 GB VRAM handles FP16 with room for concurrent users.
Best for production: RTX 5090. At 132 tok/s with 32 GB VRAM, the 5090 handles high-concurrency DeepSeek serving for production applications. Sub-4-second responses for 512-token generations.
Best budget: RTX 4060. AWQ 4-bit DeepSeek on the RTX 4060 achieves $1.14 per million tokens, 43% cheaper than the API. Good for development and cost-sensitive applications.
Best for reasoning chains: RTX 5080. DeepSeek-R1’s reasoning capabilities benefit from fast generation. At 81 tok/s the 5080 keeps multi-step reasoning chains responsive. See our AI agents GPU guide for agent-specific benchmarks.
For deployment instructions, see our DeepSeek deployment guide. For RAG integration, see best GPU for RAG pipelines and best GPU for LangChain.
Run DeepSeek on Dedicated GPU Servers
GigaGPU provides bare-metal GPU servers optimised for DeepSeek inference with vLLM and Ollama. Full data privacy, no rate limits, no per-token API fees.
Browse GPU Servers
