Not in the top 1,000 by GitHub stars

mit-han-lab/ncu-report-skill

History

Review weekly changes to SKILL.md before updating. Dates mark the weeks in which we observed each version; checks apply to the recorded commits.

View weekly changes

`SKILL.md`

Week of Nov 2, 2026 · +0 / −0 lines

Before Network or broad shell → After Network or broad shell

c7980a43cb88 → c7980a43cb88

Compare on GitHub

Week of Oct 26, 2026 · Line counts unavailable

Before Not checked yet → After Network or broad shell

First observation → c7980a43cb88

No earlier version was observed for comparison.

Compare on GitHub

View on GitHub Add to my skills

Verified commit c7980a43cb88a8734d113b22b57ff66746cf5719

274GitHub stars

7stars in 7 days

1skill in this repository

8commits in 30 days

Heat

4.8

Star growth, last 7 daysNot available
Star growth, last 24 hoursNot available
Rank movement, last 7 daysNot available
Commits, last 30 daysNot available
Accounts asking about it through MCP, last 7 daysNot available
Hot votes this weekNot available

Heat never changes the order by stars, and sponsorship never touches it. How heat works

Stars over time

30-day change not yet available: not enough history.

Individual skills in this repository

ncu-report-skill /

How it works

ncu-report-skill

When your agent uses it

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.

What it loads into context

Frontmatter

194 tokens

Always available to the agent

SKILL.md

2,841 tokens

When the skill is activated

Reference documents

48,322 tokens

Loaded on request

Token counts are estimates. Scripts run separately; they are not counted as reference context.

The playbook

Golden rule
Quickstart (what to do when someone says "profile this kernel")
Where the GPU is
File index
Critical lessons (don't skip)
Related skills

Scripts executed separately

helpers/analyze_reports.py
helpers/extract_stall_hotspots.py
helpers/list_flashinfer_workloads.py
helpers/ncu_utils.py
helpers/plot_timeline.py

Files

.gitignoreasset391 bytes
LICENSEasset1,063 bytes
README.mddoc7,283 bytes
SKILL.mdcore11,700 bytes
SKILL.mdcore11,700 bytes
SKILL.mdcore11,700 bytes
SKILL.mdcore11,700 bytes
SKILL.mdcore11,700 bytes
SKILL.mdcore11,700 bytes
SKILL.mdcore11,700 bytes
blackwell-cuda-programming.mddoc45,341 bytes
helpers/README.mddoc3,284 bytes
helpers/analyze_reports.pyruns6,037 bytes
helpers/extract_stall_hotspots.pyruns11,279 bytes
helpers/harness_template.cuasset7,801 bytes
helpers/list_flashinfer_workloads.pyruns11,386 bytes
helpers/ncu_utils.pyruns23,813 bytes
helpers/plot_timeline.pyruns8,802 bytes
helpers/safetensors_loader.hasset6,801 bytes
reference/00-directory-layout.mddoc7,930 bytes
reference/01-workflow.mddoc10,556 bytes
reference/02-harness-guide.mddoc14,370 bytes
reference/03-collection.mddoc12,423 bytes
reference/04-python-api.mddoc12,264 bytes
reference/05-analysis-dimensions.mddoc17,785 bytes
reference/06-diagnosis-playbook.mddoc17,769 bytes
reference/07-report-template.mddoc6,583 bytes
reference/08-metric-names.mddoc13,409 bytes
reference/09-common-issues.mddoc12,844 bytes

Read SKILL.md


# 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.

View the complete source on GitHub

Skills included

ncu-report-skill

Repository facts

License: MIT
Created: Apr 13, 2026
Last pushed: Oct 14, 2026
Forks: 39
Open issues: 1
Project activity: Active

Before you install

Network or broad shell

Automated signals, not a guarantee

Checked Nov 6, 2026

Not fully analyzed: some files were too long, unreadable or missing. This never lowers the level.

Show 6 notes

SKILL.md:1 Analysis incomplete: Referenced executable target unavailable; not analyzed
SKILL.md:1 Analysis incomplete: Referenced executable target unavailable; not analyzed
SKILL.md:1 Analysis incomplete: Referenced executable target unavailable; not analyzed
SKILL.md:1 Analysis incomplete: Referenced executable target unavailable; not analyzed
SKILL.md:1 Analysis incomplete: Referenced executable target unavailable; not analyzed
SKILL.md:1 Analysis incomplete: Referenced executable target unavailable; not analyzed

No broad tool permissions
9 dangerous shell commandsin 4 files: reference/09-common-issues.md:13 reference/03-collection.md:235 SKILL.md:51 and 1 more
13 network referencesin 2 files: README.md:2 reference/08-metric-names.md:142
No hidden or obfuscated text
No instruction overrides
5 executable scriptsin 5 files: helpers/analyze_reports.py:1 helpers/extract_stall_hotspots.py:1 helpers/list_flashinfer_workloads.py:1 and 2 more
1 reference to secretsin 1 file: helpers/list_flashinfer_workloads.py:46
No requests to vote in Hot

Show all 28 signals by file

`README.md`

Network accessline 2

> [!IMPORTANT] > This skill is maintained as a standalone submodule of > [Kernel Design Agents (KDA)](https://github.com/mit-han-lab/kernel-design-agents)

Network accessline 3

> This skill is maintained as a standalone submodule of > [Kernel Design Agents (KDA)](https://github.com/mit-han-lab/kernel-design-agents) > for easy installation.

Network accessline 4

> [Kernel Design Agents (KDA)](https://github.com/mit-han-lab/kernel-design-agents)

Network accessline 6

> > For bug reports, feature requests, and discussions, please use the main KDA repository: > https://github.com/mit-han-lab/kernel-design-agents

Network accessline 8

> https://github.com/mit-han-lab/kernel-design-agents

Network accessline 62

# Clone somewhere stable git clone https://github.com/mit-han-lab/ncu-report-skill.git ~/workspace/ncu-report-skill

Network accessline 63

git clone https://github.com/mit-han-lab/ncu-report-skill.git ~/workspace/ncu-report-skill

Network accessline 82

git clone https://github.com/mit-han-lab/ncu-report-skill.git /tmp/ncu

Network accessline 90

cd /path/to/other-repo git submodule add https://github.com/mit-han-lab/ncu-report-skill.git .claude/skills/ncu-report-skill

Network accessline 91

git submodule add https://github.com/mit-han-lab/ncu-report-skill.git .claude/skills/ncu-report-skill

`SKILL.md`

Dangerous shell operationline 51

ncu locks the GPU clock for every profiled kernel by default, which keeps reports comparable: at boost in ncu 2026.1–2026.3 (`--clock-control boost`; older versions default to `base`, which on a B300 runs kernels ~1.7x slower). `ncu --help`

`helpers/analyze_reports.py`

Executable scriptline 1
```
#!/usr/bin/env python3
```

`helpers/extract_stall_hotspots.py`

Executable scriptline 1
```
#!/usr/bin/env python3
```

`helpers/list_flashinfer_workloads.py`

Executable scriptline 1
```
#!/usr/bin/env python3
```
Access to secretsline 46
```
[Credential-related source omitted]
```

`helpers/ncu_utils.py`

Executable scriptline 1
```
Executable file
```

`helpers/plot_timeline.py`

Executable scriptline 1
```
#!/usr/bin/env python3
```

`reference/01-workflow.md`

Dangerous shell operationline 56

# 5. Permissions. ncu usually works without sudo (RestrictProfilingToAdminUsers is 0

`reference/03-collection.md`

Dangerous shell operationline 235
```
sudo nvidia-smi -lgc <mhz>,<mhz>
```
Dangerous shell operationline 237
```
sudo nvidia-smi -rgc
```

`reference/08-metric-names.md`

Network accessline 142

l1tex__t_requests_pipe_lsu_mem_global_op_ld.sum                    # LD request count l1tex__t_requests_pipe_lsu_mem_global_op_st.sum                    # ST request count # sectors/request = sectors.sum / requests.sum (ideal: 4 / 8 / 16 fo

Network accessline 143

l1tex__t_requests_pipe_lsu_mem_global_op_st.sum                    # ST request count # sectors/request = sectors.sum / requests.sum (ideal: 4 / 8 / 16 for 4- / 8- / 16-byte loads per thread) smsp__sass_average_data_bytes_per_sector_mem_glo

Network accessline 144

# sectors/request = sectors.sum / requests.sum (ideal: 4 / 8 / 16 for 4- / 8- / 16-byte loads per thread)

`reference/09-common-issues.md`

Dangerous shell operationline 13

**A) Use sudo (simplest on dedicated servers):**

Dangerous shell operationline 15
```
sudo ncu [...]
```

Dangerous shell operationline 20

sudo sh -c 'echo "options nvidia NVreg_RestrictProfilingToAdminUsers=0" > /etc/modprobe.d/ncu.conf'

Dangerous shell operationline 21
```
sudo update-initramfs -u
```
Dangerous shell operationline 194
```
sudo nvidia-smi -pm 1
```

How the signals fit together

Facts from a static read of the files at the pinned commit. They describe what the code does, not why.

Signal confidence: 28 facts

Network access in prose of documentationREADME.md:2Low confidence
Network access in prose of documentationREADME.md:3Low confidence
Network access in prose of documentationREADME.md:4Low confidence
Network access in prose of documentationREADME.md:6Low confidence
Network access in prose of documentationREADME.md:8Low confidence
Network access in code of documentationREADME.md:62Low confidence
Network access in code of documentationREADME.md:63Low confidence
Network access in code of documentationREADME.md:82Low confidence
Network access in code of documentationREADME.md:90Low confidence
Network access in code of documentationREADME.md:91Low confidence
Dangerous shell operation in prose of the SKILL.mdSKILL.md:51Low confidence
Executable script in code of a file that runshelpers/analyze_reports.py:1High confidence
Executable script in code of a file that runshelpers/extract_stall_hotspots.py:1High confidence
Executable script in code of a file that runshelpers/list_flashinfer_workloads.py:1High confidence
Access to secrets in code of a file that runs: a variable, placeholder or call, not a stored valuehelpers/list_flashinfer_workloads.py:46Low confidence
Executable script in prose of a file that runshelpers/ncu_utils.py:1Low confidence
Executable script in code of a file that runshelpers/plot_timeline.py:1High confidence
Dangerous shell operation in code of documentationreference/01-workflow.md:56Low confidence
Dangerous shell operation in code of documentationreference/03-collection.md:235Low confidence
Dangerous shell operation in code of documentationreference/03-collection.md:237Low confidence
Network access in code of documentationreference/08-metric-names.md:142Low confidence
Network access in code of documentationreference/08-metric-names.md:143Low confidence
Network access in code of documentationreference/08-metric-names.md:144Low confidence
Dangerous shell operation in prose of documentationreference/09-common-issues.md:13Low confidence
Dangerous shell operation in code of documentationreference/09-common-issues.md:15Low confidence
Dangerous shell operation in code of documentationreference/09-common-issues.md:20Low confidence
Dangerous shell operation in code of documentationreference/09-common-issues.md:21Low confidence
Dangerous shell operation in code of documentationreference/09-common-issues.md:194Low confidence

Works with

Agent compatibility, from the files in this repository
Agent	Status	Evidence	Install to
Claude Code	Compatible	Standard SKILL.md SKILL.md	`.claude/skills/<skill-name>/`
Codex	Compatible	Standard SKILL.md SKILL.md	`.agents/skills/<skill-name>/`
Cursor	Compatible	Standard SKILL.md SKILL.md	`.cursor/skills/<skill-name>/`
Gemini CLI	Compatible	Standard SKILL.md SKILL.md	`.gemini/skills/<skill-name>/`
GitHub Copilot	Compatible	Standard SKILL.md SKILL.md	`.github/skills/<skill-name>/`
OpenCode	Compatible	Standard SKILL.md SKILL.md	`.opencode/skills/<skill-name>/`

Native The repository ships a manifest, folder or frontmatter field for this agent.
Compatible A standard SKILL.md that this agent can read.
Unknown No evidence either way.

Derived automatically from paths, manifests and frontmatter. Not tested on each agent.

mit-han-lab/ncu-report-skill

History

SKILL.md

Share this page

Heat

Stars over time

Individual skills in this repository

How it works

ncu-report-skill

Frontmatter

SKILL.md

Reference documents

The playbook

Scripts executed separately

Files

Skills included

Repository facts

Before you install

README.md

SKILL.md

helpers/analyze_reports.py

helpers/extract_stall_hotspots.py

helpers/list_flashinfer_workloads.py

helpers/ncu_utils.py

helpers/plot_timeline.py

reference/01-workflow.md

reference/03-collection.md

reference/08-metric-names.md

reference/09-common-issues.md

How the signals fit together

Works with

You might also use

mattpocock/skills

obra/superpowers

NousResearch/hermes-agent

multica-ai/andrej-karpathy-skills

anthropics/skills

DietrichGebert/ponytail

`SKILL.md`

`README.md`

`SKILL.md`

`helpers/analyze_reports.py`

`helpers/extract_stall_hotspots.py`

`helpers/list_flashinfer_workloads.py`

`helpers/ncu_utils.py`

`helpers/plot_timeline.py`

`reference/01-workflow.md`

`reference/03-collection.md`

`reference/08-metric-names.md`

`reference/09-common-issues.md`