Agentic CUDA Kernel Optimizer
github.com/bertaye/agentic-cuda-optimizer- Category
- AI Agents
- Rank
- No. 3513Tools index
Previous survey · No. 3372 ·
- Type
- AGENT
- Use case
- Coding
- GitHub
- 37 stars
- Date
About
A LangGraph-based agent that turns a workload description into a working GPU kernel through repeated cycles of code generation, correctness checking, benchmarking, and refinement. It compiles candidate kernels with NVRTC through a standalone C++ test harness, can query GPU properties and NVIDIA documentation, inspects Nsight Compute profiling counters to guide its next attempt, and keeps the fastest implementation that passes validation.
At a glance
CUDA Kernel Optimizer generates GPU kernel candidates, checks their output against a reference and keeps the fastest one that passes. It is worth evaluating when you already have representative inputs and a trustworthy baseline.
Good fit
GPU developers who want to explore implementation and launch settings on a specific device.
Consider something else when
Build a reference and test corpus first if you cannot tell whether a faster candidate is correct.
Start here
Supply a small known-good kernel and synthetic inputs. Run a bounded experiment and independently replay the winning candidate on held-out cases.
Cost and requirements
Requires an NVIDIA GPU, compatible CUDA tools, a C++17 build environment and an OpenAI API key. Model requests are billable.
Limits to know
Candidates execute on your GPU and are only checked against the supplied cases. Run in an isolated workspace, review generated code and check the saved artifacts before sharing.
Why it made the leaderboard
Automates a tedious, expertise-heavy loop, writing and benchmarking CUDA kernel variants with real GPU profiling feedback, instead of just wrapping a chat model around code generation.
Tags
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.