SemiAnalysis podcast says GLM 5.3 writes faster GPU kernels than humans

SemiAnalysis hosts Kimbo, Wenyao and Jordan say GLM 5.3 writes faster GPU kernels than humans when given enough budget, according to an X post promoting their podcast episode.

The episode also covers sparse attention and a memory bottleneck the hosts say is not priced in. They argue that cheaper tokens could increase demand for HBM memory.

The only report so far is the SemiAnalysis post, which relays the hosts' claims without benchmark data or named kernels in the source provided.