Our take
Newless than six months old- Activity
- 35
- Community
- 49
- Issues
- 65
- Pull requests
- 60
+182 stars in the last 30 days
- Fewer contributors than 82% of the projects we track
KDA is a promising, actively developed workflow prototype for CUDA kernel optimization, but it is guidance and tooling for agent sessions rather than a turnkey optimization system.
Good fit if
- You work on CUDA kernels and can supply the task-specific evaluator, workloads, and hardware.
- You want agents to follow a documented iterative process with recorded plans, validation, profiling, and candidate decisions.
- You are comfortable adopting an early-stage project that has recent development and prompt and workflow materials, but no published releases in the past year.
Look elsewhere if
- You need a ready-to-run kernel library or a built-in benchmark harness, neither of which the project presents as its purpose.
- You need a stable released product rather than a research prototype under active development.
- Your wishlist request targets hardware other than the currently supported B200 and B300 GPUs.
All health signals
| Last commit | 2026-10-19 (10 days ago) |
|---|---|
| Commits, last 90 days | 13 |
| Releases, last 12 months | 0 |
| Contributors | 3 (top contributor: 38% of commits) |
| Issues closed, last 90 days | 3 (typically closed in 1 day) |
| Pull requests merged, last 90 days | 3 (typically merged in 0 days) |
| Project age | 6 months |
Checked on 2026-10-30 with the GitHub API.
Project history
Daily GitHub snapshots since Oct 7, 2026. History covers stars and forks; health signals are not tracked historically.
Stars
1,587Forks
138Change since Nov 16, 2026: +3 stars·0 forks
Overview
Kernel Design Agents (KDA) is an agent-centric workflow for using coding agents to research, implement, verify, and iterate on performance-sensitive CUDA kernels. It is aimed at engineers and researchers who need a repeatable way to guide kernel optimization, rather than a ready-made kernel library or a benchmark harness.
KDA keeps the workflow independent of a particular evaluator or hardware target, so teams can bring their own task definitions, workloads, profiling tools, and validation commands. Its community kernel wishlist currently accepts requests for NVIDIA B200 and B300 GPUs.
Key Features
- A documented workflow for planning, implementing, validating, and recording kernel optimization work.
- A starter prompt for turning a kernel task into an agent-assisted implementation plan.
- Guidance for tracking candidates, benchmark results, profiling evidence, and promotion decisions.
- A recommended separate workspace layout for plans, runs, outputs, profiles, and candidate records.
- Skills for NCU reports and KernelWiki, provided as pinned Git submodules.
- A community wishlist for reproducible kernel optimization requests, with submissions made as pull requests.
- Export support for validated CuTe candidates as Hugging Face kernels.
Use Cases
- CUDA engineers can use the workflow to organize agent-assisted optimization of a kernel with a defined correctness check and benchmark.
- Research teams can record implementation attempts and profiling evidence so collaborators can understand why a candidate was selected.
- Teams with a custom evaluator can adapt KDA to their own workloads without adopting a prescribed benchmark harness.
- Kernel requesters can submit reproducible optimization tasks to the community wishlist when targeting supported B200 or B300 GPUs.
What you need
Detected in the repository
- Python (main language on GitHub)
- A test suite
Getting Started
Clone the repository with its submodules, then follow the README to link the skills and set up the agent session:
git clone --recurse-submodules https://github.com/NVlabs/kda.git
See the README for setup details and the minimal workflow.
License in Plain Words
No standard license was detected. Without a license, the default copyright rules apply: you may read the code, but not reuse it. Check the repository or ask the authors.
Considerations
KDA is less than six months old and explicitly describes itself as an early research prototype. Development is recent, with commits and merged pull requests in the last 90 days, and issues have been handled quickly; however, there are no releases in the past year. The project has three contributors, with work not heavily concentrated in one person.
The repository documents a workflow, not a complete execution environment: users need to provide task-specific validation, evaluation, and hardware. Its setup depends on Claude Code and the documented skills; the KernelWiki submodule also has upstream artifact terms, with restricted CuTe DSL artifacts omitted. First-party documentation and prompt-style content use CC BY 4.0, first-party source code uses Apache 2.0, and third-party submodules have separate terms. Review the repository's license and third-party notices before redistributing components.
Found this useful?
Share it with someone who would like kda.