News for hackers. Stuff that matters.
Re:How to Run DeepSeek R1 671B Locally on a $2000 EPYC Server by bick_nyers on Saturday February 01, 2025 @01:50PM (#42901001)
It would not be that slow as it is an MoE model with 37b activated parameters.

Still, 8x3090 gives you ~2.25 bits per weight, which is not a healthy quantization. Doing bifurcation to get up to 16x3090 would be necessary for lightning fast inference with 4bit quants.

At that point though it becomes very hard to build a system due to PCIE lanes, signal integrity, the volume of space you require, the heat generated, and the power requirements.

This is the advantage of moving up to Quadro cards, half the power for 2-4x the VRAM (top end Blackwell Quadro expected to be 96GB).

Re:How to Run DeepSeek R1 671B Locally on a $2000 EPYC Server by deoxykev on Monday February 03, 2025 @12:45PM (#42920777)
Yeah, there is a clear bottleneck somewhere in llama.cpp. Even high end hardware is struggling to get good numbers. The theoretical limit should be higher, but it's not yet.

Benchmarks: https://github.com/ggerganov/llama.cpp/issues/11474#issuecom...