Claude · Codex · sankalp.bearblog.dev
Auto-research with codex: How I achieved a 232x Faster Kernel
Compiled by KHAO Editorial — aggregated from 1 source + 1 reference discovered via search. See llms.txt for citation guidance.
◌ Single Source
GPU Mode, in collab with Core Automation, recently hosted an auto-research themed contest.
Key facts
- The important sizes were batched square matrices like 512 x 512, with larger 1024, 2048, and 4096 cases too
- For the 3×3 example above, the first reflector maps the first column (12, 6, −4) straight onto (−14, 0, 0) in one shot
- Take the first column of the 3×3 example above: (12, 6, -4)
- This turns the current column into (-14, 0, 0) and rewrites the other columns consistently, so the next reflector is built from the updated matrix
Summary
Make serial work small with the help of the blocked Householder algorithm. The reporter placed 12th out of 183 participants, ending up with a 232x speedup over the baseline solution. Note that you don't need to go through the mathematics or the problem itself in detail to follow most of this blog post. You can check out the full contest page here: Problem Link and Leaderboard. This contest was part of GPU Mode's Linear Algebra Kernels in the Age of Research series.