Auto-research with codex: How I achieved a 212x Faster Kernel over baseline with Codex in GPU Mode's qr_v2 problem

342 · Sankalp · July 8, 2026, 5:24 p.m.
Summary
This blog post details the author's experience participating in an auto-research contest focused on GPU computing, specifically achieving a 212x speedup in QR decomposition using the blocked Householder algorithm and Codex. The author shares insights on their approach, challenges faced, and techniques used to maximize kernel performance, while highlighting the learning opportunities provided by the contest format. Overall, it combines technical depth with personal experiences in optimizing algorithms for GPU applications.