NVlabs

kda: Guide Coding Agents Through CUDA Kernel Optimization

KDA is a workflow for using coding agents to research, implement, validate, and improve performance-sensitive CUDA kernels. It helps teams structure experiments and preserve evidence for choosing a final implementation.

Other licenseAI AgentsCoding AgentPerformance
View on GitHub Compare
Stars
1.2k
Forks
108
Stars, last 10 days
+146

Our take

Health score

46/100

How it is scored
Activity
35
Community
49
Issues
53
Pull requests
47

+146 stars in the last 10 days

  • Fewer contributors than 82% of the projects we track

Overview

KDA provides a process for applying coding agents to CUDA kernel optimization. Rather than tying the work to one benchmark harness, it emphasizes defining a task contract, iterating against validation, and recording profiling and evaluation evidence.

It is intended for engineers and researchers who already have a concrete kernel task and want agents to assist with the optimization process. The repository is an early research prototype, not a CUDA kernel library or a turnkey performance evaluator.

Key Features

  • A minimal workflow from task definition through implementation, verification, and promotion of a candidate.
  • A starter prompt for describing a task to a coding agent.
  • Guidance for turning a draft plan into executable implementation steps.
  • A suggested workspace layout for plans, runs, profiling data, benchmarks, and candidate records.
  • A workflow designed to accommodate task-specific evaluators, datasets, tools, and hardware.
  • A community wishlist process for submitting kernel optimization requests, currently scoped to NVIDIA B200 and B300 GPUs.

Use Cases

  • CUDA developers can use the workflow to organize agent-assisted optimization of a defined kernel.
  • Research teams can keep validation results and profiling evidence alongside competing implementations.
  • Engineers evaluating coding agents can adapt the workflow to their own benchmark harness and hardware.
  • GPU workload owners can submit reproducible optimization requests through the community wishlist when their target is a supported B200 or B300 GPU.

Getting Started

Start with the repository README and the agent-flow guide. The documented workflow uses Claude Code skills and a task-specific workspace; follow the README for dependency setup and third-party license details.

Considerations

  • The project describes itself as an early research prototype and remains under active development.
  • KDA documents a workflow; it does not provide a universal evaluator or guarantee performance improvements. Each task needs its own validation and evaluation setup.
  • Using it for CUDA kernel work requires an appropriate CUDA-capable environment. The wishlist is explicitly limited to NVIDIA B200 and B300 GPUs.
  • The README lists third-party submodules with separate licensing terms. Review the notices and pinned-submodule guidance before using those components.

Found this useful?

Share it with someone who would like kda.