Won 5th Place in GPU Mode with Coding Agents, No CUDA Background

tokenbender · x · 2026-08-10

A developer participated in a GPU Mode kernel competition for batched Cholesky factorization on the B200 with very little CUDA background.

Instead of manual coding, they engineered an automated research loop around coding agents. The workflow included goal setting, rules, experiment memory, a verifier, and compute resources.

The agent-assisted solution drastically reduced execution time from a baseline of 2533.030 μs to just 265.234 μs, securing 5th place overall in the competition.

Original post →

More from coding & agent

coding & agent channel →