Inferact, AMD and NVIDIA each push LLM decoding with megakernels

SemiAnalysis reports that three teams have each advanced LLM decoding with megakernels, which combine work that normally runs in separate programs.

Inferact runs the approach on Google TPUs, AMD uses TileRT and FlyDSL, and NVIDIA has a GB300 result now live on InferenceX.

SemiAnalysis says it will examine the engineering behind each team's approach in a follow-up thread. The report does not include performance figures for any of the three.