DeepSeek R1 on 8× H200 NVL: VRAM, speed and fit
DeepSeek R1 at FP8 runs on the 8× H200 NVL with 160K of context at about 64 tokens a second.
Text generation · Chat · Ollama / LM Studio
Runs comfortably · on 8× H200 NVL · FP8 · Ollama / LM Studio · Chat
- GPU memory
- 658 GB
- of 1,128 GB
- Context
- 160K
- auto
- Speed
- 64 tok/s
- 52 when full
- First reply
- 0.3 s
- 2K prompt
Where the memory goes
- Model
- 641 GB
- Context
- 10.7 GB
- Overhead
- 5.7 GB
- OS
- 0.0 GB
- Free
- 470 GB
- Memory
- 658 GB of 1,128 GB
- Context
- 160K tokens · auto
- Speed
- 64 tok/s · 52 when full
- First reply
- 0.3 s for a 2K prompt
- Running cost
- $65/mo at home · $1,162/mo on a cloud 4× MI300X
Validation: Validated
| Check | Target | Result | Status |
|---|---|---|---|
| Fits in GPU memory | 1,128 GB usable | 658 GB needed | Pass |
| Headroom | 5% free | 470 GB free (42%) | Pass |
| Context for chat | 16K or more | 160K | Pass |
| Writing speed | 15 tok/s or more | 64 tok/s | Pass |
| First reply | under 3 s, 2K prompt | 0.3 s | Pass |
Targets: a reply you can read along with, a first reply the use case tolerates, and 5% of memory left free. Calculated from published specs, not benchmarked.
The setup
- Model
- DeepSeek R1
- Parameters
- 684B, 40.1B active
- Precision
- FP8
- Context
- 160K, auto
- Chats at once
- 1
- Use
- Chat
- Runs in
- Ollama / LM Studio
- Hardware
- 8× H200 NVL
- GPU can use
- 1,128 GB of 1,128 GB
Running cost
- At home
- $65/mo
- 2960 W at the wall
- Renting instead
- $1,162/mo
- 4× MI300X at $9.56/h
4 h a day, power at $0.18 per kWh, cloud at RunPod on-demand prices as of Nov 21, 2026.
How to run it
- llama.cpp
llama-server -hf unsloth/DeepSeek-R1-GGUF- vLLM
vllm serve deepseek-ai/DeepSeek-R1- MLX (Mac)
mlx_lm.chat --model mlx-community/DeepSeek-R1-4bit- Model card
- huggingface.co/deepseek-ai/DeepSeek-R1
DeepSeek R1 on similar machines
- 8× RTX PRO 6000 BlackwellRuns
- A100 80 GB SXMToo big
- H100 SXMToo big
- 4× H200 NVLRuns
All machines: DeepSeek R1 VRAM requirements. Everything that runs here: what can the 8× H200 NVL run?
Other models that run on 8× H200 NVL
Questions
Can DeepSeek R1 run on 8× H200 NVL?
DeepSeek R1 at FP8 runs on the 8× H200 NVL with 160K of context at about 64 tokens a second.
Report VR-0AADD3U, calculated Nov 22, 2026 from published specs, not benchmarked.