An AI agent reportedly wrote and tuned a GPU kernel from scratch to within about 13% of expert hand-written speed—in a few hours.
OpenBMB demonstrated ForgeKernel live: an agent that repeatedly writes and tunes a DeepSeek GPU kernel. The reported result offers a concrete AI coding use case with a performance target beyond simply producing code.
Why kernels matter
A GPU kernel is code that performs a specific piece of work on a graphics processor. Tuning it means refining how that work runs, with performance measured against a reference—in this demonstration, an expert-written version.
Before using the approach
A useful evaluation would check that the generated kernel produces correct results, repeat the timing comparison under matching conditions, and record the resources spent tuning it.
For teams considering this approach, the practical next step is a controlled trial on a relevant kernel, with correctness, repeatable performance, and tuning cost as the acceptance criteria.