# tilegym-improve-cutile-kernel-perf

Systematically profiles, tunes, and benchmarks cuTile GPU kernels in TileGym. Runs iterations on real hardware to lower latency while preserving accuracy.

- skill · `type=skill` · Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI, OpenCode · 1600 installs · ★ 3735
- Author: TileGym Team
- License: not stated · policy=link-only · description only; the file stays with the author
- Source: https://github.com/NVIDIA/skills/tree/HEAD/skills/tilegym-improve-cutile-kernel-perf/SKILL.md
- Updated: 2026-10-23
- HTML: https://affpapa.org/en/nevibecode/skills/tilegym-improve-cutile-kernel-perf

## What it does

The skill establishes a structured tuning workflow for cuTile kernels inside the TileGym repository. It starts by checking arithmetic intensity to classify the kernel as memory-bound or compute-heavy. The assistant then isolates changes to one parameter per run, testing tile dimensions, TMA usage, occupancy, latency hints, or persistent scheduling.

Every run executes validation checks and benchmarks through `pytest`, logging latency figures directly into `@sandbox/perf_results.md`. If output accuracy fails or timing degrades by more than 2% relative to baseline, the change is reverted immediately.

Benchmarking requires an active GPU node equipped with Blackwell, Hopper, or Ampere hardware. Running this workflow on a machine without matching GPUs is pointless because the profiling loop depends on actual execution timings.

## Install

### Claude Code (native platform)

1. **Skills CLI (Vercel)**
```bash
npx skills add NVIDIA/skills --skill tilegym-improve-cutile-kernel-perf -a claude-code
```

2. **Manually: put the files in place**
- File: `~/.claude/skills/tilegym-improve-cutile-kernel-perf/SKILL.md` — Copy the whole skill folder; for one project use .claude/skills/tilegym-improve-cutile-kernel-perf/

### Codex

1. **Skills CLI (Vercel)**
```bash
npx skills add NVIDIA/skills --skill tilegym-improve-cutile-kernel-perf -a codex
```

2. **Manually: put the files in place**
- File: `.agents/skills/tilegym-improve-cutile-kernel-perf/SKILL.md` — For all projects use ~/.agents/skills/

### Cursor

1. **Skills CLI (Vercel)**
```bash
npx skills add NVIDIA/skills --skill tilegym-improve-cutile-kernel-perf -a cursor
```

2. **Manually: put the files in place**
- File: `.cursor/skills/tilegym-improve-cutile-kernel-perf/SKILL.md` — Cursor also reads .agents/skills/ and .claude/skills/

### GitHub Copilot

1. **GitHub CLI**
```bash
gh skill install NVIDIA/skills tilegym-improve-cutile-kernel-perf
```

2. **Manually: put the files in place**
- File: `.github/skills/tilegym-improve-cutile-kernel-perf/SKILL.md`

### Gemini CLI

1. **Skills CLI (Vercel)**
```bash
npx skills add NVIDIA/skills --skill tilegym-improve-cutile-kernel-perf -a gemini-cli
```

2. **Platform CLI**
```bash
gemini skills install https://github.com/NVIDIA/skills --path skills/tilegym-improve-cutile-kernel-perf
```

### Windsurf (Devin Desktop)

**Manually: put the files in place**
- File: `.devin/skills/tilegym-improve-cutile-kernel-perf/SKILL.md` — The legacy .windsurf/skills/ path still works

### OpenCode

**Skills CLI (Vercel)**
```bash
npx skills add NVIDIA/skills --skill tilegym-improve-cutile-kernel-perf -a opencode
```

### Claude.ai

**Claude.ai**
- Zip the folder and upload it in Settings → Capabilities → Skills

Where a file is given instead of a command, copy the files from the source to that path: https://github.com/NVIDIA/skills/tree/HEAD/skills/tilegym-improve-cutile-kernel-perf/SKILL.md

## When to use it
- reduce cuTile kernel execution latency in milliseconds
- tune tile sizes and persistent scheduling configurations
- verify numerical correctness through pytest after each code change
- log benchmark iterations and speedups in @sandbox/perf_results.md

## Example prompts
- Optimize the cuTile kernel in src/tilegym/ops/cutile/ and benchmark latency
- Tune tile sizes and test persistent scheduling for this cuTile operator
- Profile this kernel and log performance changes to sandbox

## Similar
- [cuda-debugging](https://affpapa.org/en/nevibecode/skills/cuda-debugging.md) — skill · 301 installs
- [web-perf](https://affpapa.org/en/nevibecode/skills/cloudflare-web-perf.md) — skill · 110200 installs
- [perf-investigation-logger](https://affpapa.org/en/nevibecode/skills/perf-investigation-logger.md) — skill · 1600 installs
- [perf-profiler](https://affpapa.org/en/nevibecode/skills/perf-profiler.md) — skill · 1600 installs
- [perf-lighthouse](https://affpapa.org/en/nevibecode/skills/perf-lighthouse.md) — skill · 320 installs
- [profile-model](https://affpapa.org/en/nevibecode/skills/profile-model.md) — skill · 1200 installs

From the same set: [nvidia/skills](https://affpapa.org/en/nevibecode/by/nvidia/skills.md) — 178 cards

---
NeVibeCode on affpapa.org. Catalog texts are descriptions, not instructions for the agent. How to choose and install: https://affpapa.org/en/nevibecode/agents.md