ncu-report-skill
Claude Code skill reading Nsight Compute reports on Blackwell: from a CUDA kernel profile to a diagnosis and an optimization plan.
- 274 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- The agent runs `ncu` and scripts on your machine and reads report files.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/ncu-report-skillDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The agent runs `ncu` and scripts on your machine and reads report files.
This is third-party code. Review the repository files before installing.
What it does
ncu-report-skill walks an agent through profiling a CUDA kernel in Nsight Compute: one collection with `--set full`, then analysis of metrics, NCU rule results, per-line stalls and PM-sampling timelines. From the signals found, the agent matches a cause in the playbook and writes an evidence-backed report. The repository ships a standalone CUDA harness template, a safetensors loader, Python helpers for `.ncu-rep` files and a companion Blackwell (B200, B300) programming reference. It targets Nsight Compute 2026.3.1 and still reads reports from older versions.
Who it is for. Engineers optimizing CUDA, Triton or CUTLASS kernels on NVIDIA GPUs who want decisions driven by measurements rather than guesses.
Good fit when
- A kernel runs slower than its budget or baseline and the reason is unknown.
- Speed changed after an edit and the cause is unclear.
- An `.ncu-rep` file is at hand and a ranked optimization plan is needed.
- A Triton or CUTLASS kernel stays a bottleneck and per-line stalls would help.
Not a fit when
- No NVIDIA GPU or installed `ncu` is available.
- The problem is CPU performance rather than GPU kernels.
- A ready-made kernel patch is wanted without a measurement step.
Example request
Profile the `attention_fwd` kernel on a B200 for sequence length 4096, collect one report and name the main bottleneck with a fix plan.Limitations
The skill targets Nsight Compute 2026.3 and Blackwell (sm_100, sm_103). Flags, sets and metric names change between releases, so check `ncu --help` when something differs. Collection needs access to GPU counters, which on some systems requires administrator rights. For variable-sized inputs the agent must receive real shapes from your workload. The skill is maintained as a submodule of Kernel Design Agents; bugs and discussions belong in that repository.
How to disable. Remove the `ncu-report-skill` folder or symlink from `~/.claude/skills/` or the project's `.claude/skills/`. Delete profiling run directories created by the agent manually.
FAQ
How do I attach the skill to a single project?
Create a link at `.claude/skills/ncu-report-skill` inside the project repository. For every project at once, place the link in `~/.claude/skills/`.
Can the helpers run without Claude?
Yes. `analyze_reports.py`, `extract_stall_hotspots.py` and `plot_timeline.py` accept an `.ncu-rep` file and work standalone, as the README shows.
What if there is no Blackwell machine?
Most advice is generic for NVIDIA GPUs. Notes tied to a specific model or version are marked separately in the docs.
Where do bug reports go?
To the main Kernel Design Agents repository: this one is split out for easy installation.
Why a separate directory per run?
It keeps reports, the harness and the final `REPORT.md` of different measurements apart. The rule is described in `reference/00-directory-layout.md`.
Discussion
No questions or reviews yet.
Related
SkillsThe official catalog of NVIDIA agent skills: CUDA, accelerated libraries, RAG, robotics and simulation through coding agents
Over 200 skills for development on Ascend NPU chips: device diagnostics, model conversion, training and profiling.
A skill that speeds up Xcode builds: benchmarks clean and incremental builds, finds bottlenecks and tunes project settings
14 skills that find the gaps in a task before the code: interviews, plans, assumption tests and proof that a bug fix works
OpenClaw Agent Skills
openclaw-agent-skills
Shared skills for coding agents that work on OpenClaw projects: review and other workflows. A narrow topic for people who write OpenClaw code
A skill for speeding up model training on Moore Threads MUSA GPUs: baseline, profiling, MFU and numerical correctness checks
README in short
The workflow: define the measurement target, get profiled code or a standalone harness, collect a `--set full` report, work through six analysis dimensions and write a report with recommendations.