World's fastest panel QR factorization on B200: how a GPU MODE contestant cracked chained dependencies
A_K_Nain · x · 2026-09-27
gau-nernst took 2nd place in GPU MODE's QR factorization kernel leaderboard on B200, claims the fastest panel QR kernel by a large margin, and open-sourced the code (gau-nernst/gpu-mode-kernels).
The core challenge is chained dependency: the Householder reflector for column n depends on all previous columns, so the outer loop can't be parallelized. His approach:
- One warp computes the Householder (column reduction wants a single warp to avoid cross-warp communication)
- Trailing-column updates are independent and run across multiple warps in parallel
The competition covered 7+ shapes (batch 20–640, n up to 2048) plus rank-deficient and clustered special cases. A refreshing deep dive into non-LLM GPU kernel engineering — the author learned QR decomposition largely from Codex during the contest.
More from coding & agent
- FuSheng: AI now navigates human interfaces like a person and maintains real relationships — FuSheng_0306 · 2026-09-27
- AI agent sends personalized Mid-Autumn greetings to 50 contacts by seeing the screen and clicking, then replies to 70+ messages — FuSheng_0306 · 2026-09-27
- FuSheng had GPT-6 take over his WeChat and send 120+ error-free holiday greetings — FuSheng_0306 · 2026-09-27
- Pairing Astra and Opus for coding: 3-stage pipeline tops at 82.8, but solo runs cost far less — kevinkern · 2026-09-27
- UT Arlington paper: AI agents miss requirements even when marking tasks complete — alex_verem · 2026-09-27
- KeySync puts your 2FA codes in the Chrome toolbar, copying in 1.2s — aliscodes · 2026-09-27