DeepSeek-R1 671B (MoE) on NVIDIA A100 80GB SXM (80GB VRAM)
Even in 4-bit Q4_K_M, DeepSeek-R1 671B (MoE) requires 420.2 GB VRAM (6x NVIDIA A100 80GB SXM GPUs needed for full VRAM execution, or heavy CPU RAM offloading at ~7.3 tok/s).
🎛️ Interactive VRAM & KV-Cache Tuner
📊 Real-Time Memory Footprint
420.2 GB / 80 GB🚀 Ready-to-Run Launch Command
ollama run deepseek-r1:671b # Or with llama.cpp server (8K context, Flash Attention): llama-server -hf deepseek-ai/DeepSeek-R1-GGUF -c 8192 -ngl 99 -fa
Need 420.2 GB+ VRAM to run DeepSeek-R1 671B (MoE)?
Recommended cloud setup for 4-bit: Multi-GPU H100 / H200 Node (or 1383.46 GB for full FP16). Spin up a pre-configured vLLM / Ollama instance in 60 seconds:
| Provider | Best For | RTX 4090 (24GB) | A100 (80GB) | H100 (80GB) | Bonus / Trial | Action |
|---|---|---|---|---|---|---|
| RunPod | 1-Click vLLM / Ollama Pods | $0.44 / hr | $1.39 / hr | $2.39 / hr | Instant PyTorch / vLLM Pods + Serverless GPU | Deploy on RunPod → |
| Vast.ai | Cheapest Spot GPU Market | $0.29 / hr | $0.95 / hr | $1.79 / hr | Save 50–70% on RTX 4090 / 5090 / A100 rentals | Check Spot GPUs → |
| Vultr Cloud GPU | Enterprise SLA & Bare Metal | — | $1.79 / hr | $2.69 / hr | $300 Free Credit for New Users | Claim $300 Credit → |
| DigitalOcean GPU Droplets | Managed H100 Droplets | — | — | $2.99 / hr | $200 Free Trial (60 Days) | Get $200 Free → |
All Quantization Formats: DeepSeek-R1 671B (MoE) on NVIDIA A100 80GB SXM
| Quantization Format | Bits/Weight | Weights VRAM | KV Cache (8K) | Total VRAM | Fits on 1x NVIDIA A100 80GB SXM (80GB)? | Est. Speed | GPUs Needed |
|---|---|---|---|---|---|---|---|
| FP16 / BF16 (Unquantized) | 16 bpw | 1382.26 GB | 0.54 GB | 1383.46 GB | No (Over by 1303.5GB) | ~2.2 tok/s | 18x NVIDIA A100 80GB SXM |
| FP8 (W8A8 / E4M3) | 8.5 bpw | 734.33 GB | 0.54 GB | 735.52 GB | No (Over by 655.5GB) | ~4.1 tok/s | 10x NVIDIA A100 80GB SXM |
| GGUF Q8_0 (8-bit) | 8.5 bpw | 734.33 GB | 0.54 GB | 735.52 GB | No (Over by 655.5GB) | ~4.1 tok/s | 10x NVIDIA A100 80GB SXM |
| GGUF Q6_K (6-bit) | 6.56 bpw | 566.73 GB | 0.54 GB | 567.93 GB | No (Over by 487.9GB) | ~5.4 tok/s | 8x NVIDIA A100 80GB SXM |
| GGUF Q5_K_M (5-bit) | 5.69 bpw | 491.57 GB | 0.54 GB | 492.77 GB | No (Over by 412.8GB) | ~6.2 tok/s | 7x NVIDIA A100 80GB SXM |
| GGUF Q4_K_M / AWQ (4-bit) | 4.85 bpw | 419 GB | 0.54 GB | 420.2 GB | No (Over by 340.2GB) | ~7.3 tok/s | 6x NVIDIA A100 80GB SXM |
| GGUF IQ4_XS (4.25 bpw) | 4.25 bpw | 367.16 GB | 0.54 GB | 368.36 GB | No (Over by 288.4GB) | ~8.3 tok/s | 5x NVIDIA A100 80GB SXM |
| GGUF Q3_K_M (3-bit) | 3.91 bpw | 337.79 GB | 0.54 GB | 338.99 GB | No (Over by 259.0GB) | ~9 tok/s | 5x NVIDIA A100 80GB SXM |
Frequently Asked Questions (DeepSeek-R1 671B (MoE) + NVIDIA A100 80GB SXM)
Can NVIDIA A100 80GB SXM (80GB) run DeepSeek-R1 671B (MoE) locally?
No — A single NVIDIA A100 80GB SXM (80GB) cannot fit DeepSeek-R1 671B (MoE) in VRAM. Even in 4-bit Q4_K_M, DeepSeek-R1 671B (MoE) requires 420.2 GB VRAM (6x NVIDIA A100 80GB SXM GPUs needed for full VRAM execution, or heavy CPU RAM offloading at ~7.3 tok/s).
How is KV Cache calculated for DeepSeek-R1 671B (MoE)?
DeepSeek-R1 671B (MoE) uses MLA-MoE with 61 transformer layers, 128 attention heads, and 8 KV heads (head dimension 128). At 8,192 tokens in FP16, KV cache consumes 0.54 GB. Enabling 8-bit KV cache (-ctk q8_0 -ctv q8_0 in llama.cpp or --kv-cache-dtype fp8 in vLLM) cuts this in half to 0.27 GB.