This repository provides a structured, searchable knowledge base for NVIDIA Blackwell (SM100) and Hopper (SM90) GPU kernel optimization, packaged as a Claude Code skill. It targets engineers and researchers seeking to understand and implement high-performance GPU kernels, offering synthesized insights and powerful query tools to accelerate development and debugging.

How It Works

KernelWiki employs a three-layer architecture: raw sources/ (PRs, blogs, docs), synthesized wiki/ pages with YAML frontmatter, and auto-generated queries/ indices. Data is derived from upstream sources, cross-referenced, and made searchable via Python scripts. Integration as a Claude Code Skill is seamless; cloning the repository into ~/.claude/skills/ auto-registers it.

Quick Start & Requirements

• Install: Clone the repository to ~/.claude/skills/KernelWiki and install dependencies:

  git clone git@github.com:mit-han-lab/KernelWiki.git ~/.claude/skills/KernelWiki
  pip install -r ~/.claude/skills/KernelWiki/requirements.txt
  

• Prerequisites: Python 3, Claude Code environment.
• Docs: Companion documentation is available within the repository (SKILL.md, references/primer.md, references/schema.md, references/examples.md).

Highlighted Details

• Comprehensive knowledge base for NVIDIA Blackwell and Hopper GPU kernel optimization techniques.
• Integrated query tools (query.py, get_page.py, grep_wiki.py) for keyword, filter, and alias-aware searching.
• Hybrid version-claim registry (data/version-claims.yaml) tracks version-sensitive claims for tools like Triton and CUTLASS.
• Rigorous quality gates ensure schema validation, link integrity, and evidence-based verification for all wiki pages.
• Verbatim/derived asset bundles under artifacts/ are pinned to upstream SHAs via PROVENANCE.yaml.

Maintenance & Community

The repository was last updated on June 9, 2026. For bug reports, feature requests, and discussions, users should refer to the main Kernel Design Agents (KDA) repository: https://github.com/mit-han-lab/kernel-design-agents.

Licensing & Compatibility

The tooling (scripts, references, data) is provided under an MIT-style license. Content is presented as derivative works citing upstream sources, implying compatibility with the licenses of the original works.

Limitations & Caveats

Information is current only up to the last repository update (June 9, 2026). The scope is strictly limited to Blackwell-first, kernel-only topics, and English canonical content; distributed systems topics are out of scope. SM90 content requires an explicit blackwell_relevance field.