09/02/2025 6:40 am
Topic starter
Notifications
Clear all
0
I planning to build a PC to run DeepSeek R1 70b locally, how much VRAM is required? And which GPU is best for this model?
3 Answers
0
10/02/2025 2:18 am
48GB, recommended GPUs: RTX 5090 x2, A100, H100.
0
11/02/2025 11:48 am
From personal experience: Running DeepSeek R1 70B with 4-bit quantization on a single RTX 6000 Ada (48GB)works for inference at batch size 1. Larger batches or training require splitting the model across GPUs. For reference, LLaMA 70B in 4-bit uses ~40GB VRAM, so expect similar here. If you’re stuck with 24GB cards (e.g., 3090/4090), you’ll need 2-3 GPUs and tensor parallelism. Always monitor VRAM usage with nvidia-smi!
0
12/02/2025 8:09 am
50GB and more.
