Skip to content
Trending storyDeveloping

Agent swarm writes 10,000 lines of fused Hopper decode kernels, up to 38% faster

1 article1 sourcesince Oct 31Last article 1h ago ·

Overview

AISummary of 1 article

mgoin used an agent swarm to write 10,000 lines of ultra-fused kernels for NVIDIA Hopper GPUs that improve decode speed by up to 38%, according to a post from SemiAnalysis.

The post says NVIDIA's decode-specialized software has focused mainly on Blackwell Ultra and Rubin, with little support for Hopper or Blackwell.

The report comes from a single X post by SemiAnalysis. It does not include benchmark details, the models tested, or the hardware configurations behind the 38% figure.

Written by AI from the articles below · updated Oct 31, 7:07 PM ET

Check the sources:

Latest developmentOct 31, 7:00 PM ET
Agent swarm writes Hopper decode kernels that speed up decoding up to 38%

Article timeline

The articles in this story. Times are ET.

Oct 31
  1. SemiAnalysisX
    Agent swarm writes Hopper decode kernels that speed up decoding up to 38%

    AImgoin used an agent swarm to write 10,000 lines of fused Hopper kernels that improve decode speeds by up to 38%. The post says NVIDIA's decode-specialized software has focused on Blackwell Ultra and Rubin rather than Hopper or Blackwell.

    Image from @SemiAnalysis_'s post

Heat Trend

Not enough continuous observations to show a trend yet.