gg2
Auto-research with codex: 232x faster kernel
Archived — this story has rotated out of today’s deck. It is kept here in full.
The gist
A developer used Codex to achieve a 232x faster GPU kernel in a contest.
It shows AI can auto-research and optimize code, potentially transforming developer workflows.
Background
GPU Mode and Core Automation hosted an auto-research contest where participants had to implement batched square compact-Householder QR factorization. The developer, Sankalp, used Codex to iteratively optimize a CUDA kernel, achieving a 232x speedup over the baseline and placing 12th out of 183 participants. The contest provided an agent-friendly environment with tools like popcorn CLI for testing and benchmarking.
How it unfolded
- recentlyGPU Mode and Core Automation hosted an auto-research contest for implementing batched square compact-Householder QR factorization.
- recentlySankalp used Codex to iteratively optimize the kernel, achieving a 232x speedup over baseline.
- recentlySankalp placed 12th out of 183 participants in the contest.
Who’s saying what
- Developer
- Sankalp described the process as 'auto-research' and acknowledged some might call it 'loop engineering', but found it effective.
- Skeptics
- Some people might dismiss this as 'loop engineering' rather than genuine research, but the developer says that is fine.