skills.rest
Romaosir
RomaosirCOMMUNITY
Explore all skills by this author →
GitHub

ncu-cuda-profiling

Automate Nsight Compute profiling and classify CUDA kernel bottlenecks.

3|Updated May 14, 2026
Search Tags:#automation#performance optimization#profiling#ncu#cuda#roofline#kernel bottleneck
One-click install
npx skills add https://github.com/Romaosir/IF_Romao_kernel_optimize --skill ncu-cuda-profiling-romaosir
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: ncu-cuda-profiling
Source: https://github.com/Romaosir/IF_Romao_kernel_optimize/tree/main/skills/b200_skill/ncu-cuda-profiling-skill
Command: npx skills add https://github.com/Romaosir/IF_Romao_kernel_optimize --skill ncu-cuda-profiling-romaosir

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes scripts (resource) and references (resource) and assets (resource) components.

What problem does it solve?

It helps you quickly identify why a CUDA kernel is slow by automating Nsight Compute (NCU) profiling, extracting key metrics, and generating actionable bottleneck diagnoses instead of manually inspecting raw profiler outputs.

Core Features & Use Cases

  • Automated full-metric NCU profiling: runs NCU with full metric collection and saves results for later reuse.
  • Metric export for analysis: imports existing .ncu-rep reports and exports raw metrics (e.g., CSV) and per-kernel summaries.
  • AI-style bottleneck diagnosis and optimization guidance: classifies common bottleneck types (e.g., DRAM/L1 pressure/latency/compute/occupancy) and recommends concrete optimization directions, including shared memory tiling/padding and other tuning approaches.
  • Use Case: When you suspect your matmul kernel is memory- or latency-bound, generate a report once, then ask the skill to diagnose the bottleneck and outline next optimization steps without repeating profiling from scratch.

Quick Start

Ask the AI to analyze a previously collected file by running an NCU import on your report, e.g., “Analyze my_report.ncu-rep and summarize the main bottleneck with optimization suggestions.”

Frequently Asked Questions about ncu-cuda-profiling

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I diagnose CUDA kernel performance bottlenecks from an Nsight Compute report?▼

To diagnose CUDA kernel performance bottlenecks, import an existing .ncu-rep file to extract key metrics and generate actionable bottleneck classifications like DRAM or compute pressure without manually inspecting raw profiler outputs.

Can I reuse saved Nsight Compute reports for repeatable kernel bottleneck analysis?▼

Yes, you can reuse saved .ncu-rep reports for repeatable kernel bottleneck analysis by importing them with ncu --import to export CSV summaries and per-kernel throughput metrics for later comparisons.

What is the best way to identify if a slow matmul kernel is memory-bound or latency-bound?▼

The best way to identify if a matmul kernel is memory-bound or latency-bound is by running automated NCU profiling with full metrics, then applying rule-based classification on extracted throughput and occupancy indicators.

Does this approach require collecting full metrics with ncu --set full for accurate profiling?▼

Yes, accurate automated bottleneck diagnosis requires full metrics collection via ncu --set full to ensure all necessary throughput and occupancy indicators are available for rule-based bottleneck classification.

How do I get optimization suggestions for CUDA kernels suffering from L1 cache pressure?▼

To get optimization suggestions for CUDA kernels with L1 cache pressure, import your .ncu-rep report to classify the bottleneck type, which then recommends concrete tuning approaches like shared memory tiling or padding.

Related Skills

Explore More Software Engineering Skills→
yo-stevenyo-steven

perf-nsight-compute-analysis

Classify SOL% bottlenecks and diagnose CUDA kernel underperformance from Nsight Compute output.

Community
Advanced
RomaosirRomaosir
3

cuda-kernel-techniques

Identify CUDA kernel optimization techniques by profiler symptoms and hardware context.

Community
Intermediate
fsx950223fsx950223
2

kernel-trace-analysis

Profile GPU kernels with rocprofv3 and generate optimization plans.

Community
Advanced
mohitmishra786mohitmishra786
159

intel-vtune-amd-uprof

Analyze CPU microarchitecture and performance bottlenecks with Intel VTune Profiler and AMD uProf.

Community
Advanced