Menu

Individual skill

ncu-report-skill

Use before deciding how to optimize a CUDA kernel (hand-written CUDA, Triton, CUTLASS/CuTe), when a kernel is slower than expected, and whenever you are about to run Nsight Compute (ncu) or read its output — instead of guessing at the bottleneck, writing your own profiling script, or reading ncu's console tables by eye. One ncu collection, then helpers that extract key metrics, NCU's rule results, per-line stalls (inlined CUTLASS/CuTe/Triton code included) and PM timelines, and a playbook from signals to fixes. Also when a user asks to profile a kernel, read an ncu report or write an optimization plan ("profile 一下", "为什么慢", "ncu 报告"). Written for ncu 2026.3 on Blackwell (B200 / B300) and other NVIDIA GPUs.

SKILL.md

Read pinned source

Source commit c7980a43cb88a8734d113b22b57ff66746cf5719

What was checked

Matches the observed repository HEAD
Yes
Scanner rules are current
No
Repository scan complete
No
Skill scan complete
No
Observed behavior
Not checked yet
Last checked
Nov 6, 2026

These are observations at the recorded commit, not a guarantee of safety. Review the source before running it.

Install this skill

Choose an agent in your Stack or copy a command below. Commands fetch the pinned commit and copy only this skill directory. Existing destinations are refused. Copying a command does not install anything.

An install command is unavailable: current rules, complete repository and skill scans, an eligible behavior level and a permitted license are required.

Repository context

Repository stars (not skill ratings)
275
Repository last updated
Oct 14, 2026
License
MIT

Observed agent integration: No specific integration evidence. Standard SKILL.md installation does not imply verified compatibility.

Files and excerpt

File metadata and permitted, redacted excerpts come from the recorded scan. Source links are pinned to that commit.

Browse files
Read excerpt

# Skill: CUDA Kernel Profiling (Nsight Compute)

**When to use:** before choosing what to optimize in a GPU kernel (CUDA C++, Triton, CUTLASS/CuTe), when a kernel is slower than its budget or baseline, after a change whose effect you do not understand, and every time you are about to run `ncu` or read a report — so the bottleneck comes from measurements, not guesses. Also when someone asks to profile a kernel, find its bottlenecks, read an ncu report or write an optimization plan ("profile X", "为什么这个 kernel 慢", "ncu report 说...", "下一步怎么优化", "帮我看一下这份 ncu 报告").

**Written for:** Nsight Compute 2026.3.1, the newest release (it also reads reports of older versions), on Blackwell — B200 (sm_100) and B300 (sm_103), 148 SMs each. A CUDA toolkit bundles the ncu of its release (CUDA 13.2: 2026.1.1); newer ones are separate installs (NVIDIA's CUDA package repositories carry `nsight-compute-<version>`, and the toolkit's `ncu` launcher runs the newest installed). Most advice is generic; GPU- or version-specific notes say so. ncu's flags, sets and metric names change between releases: `ncu --version`, `ncu --help`, `ncu --list-sets` and `ncu --list-sections` are the authority when something here does not match what you have.

---

## Golden rule

**Profile → Diagnose → Plan, in that order. Never guess.**

Most under-performing CUDA kernels are under-performing for exactly one reason that ncu can tell you in 10 seconds. Don't invent hypotheses before you have the report. Don't start coding a fix before you've matched the observed pattern to a known diagnosis. Don't write a wall of suggestions — rank them by evidence and expected impact.