Sankalp drives Codex to a 232x faster QR kernel on B200s

Sankalp ran GPT-5.5 Codex through more than 1,500 submissions over 14 days on GPU Mode's batched compact-Householder QR problem, cutting B200 runtime from about 419,000 microseconds to 1,805 — 232x over baseline, and 12th of 183 entrants. The transferable piece is the harness, not the kernel: keep three to five parallel candidate families alive instead of one improvement path, profile through Modal and NVIDIA's Nsight Compute, and select on measured runtime rather than reviewer judgment. The loop was steered every two to three hours, and the number is self-reported and not independently reproduced.

Source: sankalp.bearblog.dev ↗