Agentic CUDA Optimizer puts an LLM in a GPU tuning loop
The open-source tool, posted to Show HN, has an LLM propose, compile and benchmark CUDA kernel changes in a loop, keeping only versions that pass validation.
A developer known as bertaye posted Agentic CUDA Optimizer to Show HN on Friday, an open-source tool that automates the trial-and-error of tuning GPU kernels for speed.
The tool uses LangGraph to orchestrate the loop. An LLM proposes kernel and launch-parameter changes. A C++ harness compiles them with NVRTC and checks correctness against a reference implementation. A benchmarking step times each candidate with CUDA events, then feeds the results back for another pass.
On an RTX 3060 Laptop GPU, the system optimized float32 GEMM kernels and produced heatmaps tracking timing improvements across iterations, according to the project's README. The repo has drawn 23 stars on GitHub and 29 points on Show HN, where six commenters weighed in, but no outside benchmark has verified the results.
The project's own README admits the limits. Passing the supplied test cases "does not prove general correctness, and a generated reference is not an independent correctness oracle," it says. The tool needs an OpenAI API key, a CUDA Toolkit and an Nvidia GPU to run, and its author calls the project experimental.