Show HN: Agentic CUDA Kernel Optimizer
An open-source LangGraph-powered agent automates CUDA kernel generation, Nsight profiling, correctness checks, and benchmarking to optimize GPU performance.
summary is enough The repository README explains the LangGraph workflow clearly, but general coding agents can achieve similar results without a constrained framework.
summary
An open-source tool uses LangGraph to automate the optimization cycle for CUDA kernels, combining code generation, NVRTC compilation, and benchmarking. It supports optional Nsight Compute profiling and NVIDIA documentation research to guide iterative code improvements. Correctness is verified by comparing outputs against a reference kernel using a standalone C++ test harness. Commenters note that general coding agents can perform similar tasks, leading to discussions about the benefits of structured agent frameworks versus open-ended tools.
key points
- The workflow uses LangGraph to bound the optimization loop, handling code generation, compilation via NVRTC, and timing measurement.
- Optional flags enable Nsight Compute profiling and NVIDIA documentation retrieval to feed concrete performance metrics back into the model.
- Correctness is validated by comparing candidate outputs byte-by-byte against a reference kernel across specified input cases.
who should read it
Systems engineers and GPU programmers interested in AI-driven performance optimization.
discussion mixed
consensus
Commenters question whether specialized agent frameworks outperform general-purpose coding agents that can interact directly with developer tools and terminals.
pushback
Concerns were raised about correctness validation, noting that AI-generated reference tests might be gamed or miss subtle synchronization bugs.
Yup, this thing is basically useless. I just did /goal optimize the cuda kernel, and ai agent just proposed and tested a bunch of ideas by itself, and it profiled them using nsight ncu etc. by itself, which lead to one order of magnitude faster kernel.
Demonstrates that general-purpose terminal agents can achieve the same CUDA optimization results without a specialized framework.
Generated by an LLM from the article and thread. Check the sources before relying on it.