Skip to content
View original post on X: wh
Published
18/100

DeepSeek still avoids tensor parallelism even on Blackwell GPUs

AIwh says DeepSeek still does not use tensor parallelism in its inference setup, even on Blackwell GPUs with standard interconnect, a choice that traces back to DeepSeek V3 on H800s with limited interconnect. wh notes that the default recipes in vLLM and SGLang both recommend tensor parallelism.
whVerified on X
@nrehiew_
View on X

Interesting that even with Blackwells (and presumably standard interconnect), DeepSeek still don’t use any Tensor Parallelism. This traces back all the way to DeepSeek v3 on H800s with limited interconnect.

Both VLLM and Sglang’s default recipes all recommend TP

elie@eliebakouch
this might also be a confirmation that deepseek is using nvidia blackwell and NVL72 for part of their production inference?
View quoted post on X
Source details

wh · x.comPublished

Original title: Interesting that even with Blackwells (and presumably standard interconnect), DeepSeek still don’t use any Tensor Parallelism. This trace...

Topics

DeepSeekDeployment & engineeringChina AIExpert opinionTrends