K3 hit 77% speedup optimizing mjwarp kernels; GPT-6 could only add 0.38%

YouJiacheng · x · 2026-09-10

The author had K3 optimize mjwarp kernels for a workload last week, achieving a 77% speedup. Asking GPT-6 Astra to optimize further, it first failed on an obviously wrong direction, then tuned numthreads for only a 0.38% gain.

He notes GPT-6 performs well on other tasks, and that he once proposed a similar idea to K3 that was rejected via simple memory-access analysis.

Related event: K3 delivers 77% kernel speedup, GPT-6 adds little(2 posts)→

Original post →

More from coding & agent

coding & agent channel →