cuda

Analyze CUDA kernel performance bottlenecks with Nsight Systems and Nsight Compute.

1|Updated Apr 4, 2026
One-click install
npx skills add https://github.com/m0at/claudemd --skill cuda-m0at
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: cuda
Source: https://github.com/m0at/claudemd/tree/main/skills/ptx-isa
Command: npx skills add https://github.com/m0at/claudemd --skill cuda-m0at

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

CUDA developers need a structured, reliable guide to debug, profile, and optimize GPU kernels across tooling ecosystems, which can be scattered across docs and tools.

Core Features & Use Cases

  • Provides a consolidated overview of CUDA kernel development workflows, including profiling, debugging, and binary inspection.
  • Helps teams standardize performance optimization using Nsight Systems (nsys), Nsight Compute (ncu), cuda-gdb, compute-sanitizer, and cuobjdump workflows.
  • Use case: a dev investigating a slow kernel can follow the workflow to reproduce, profile, and optimize with targeted tooling steps.

Quick Start

Launch a small CUDA kernel and profile it with nsys and ncu to observe kernel timings.

Frequently Asked Questions about cuda

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

FAQPage Schema
How do I profile a CUDA kernel to identify performance bottlenecks?

Profile CUDA kernel performance bottlenecks by applying Nsight Systems and Nsight Compute to capture kernel timings and metrics. This workflow guides you through reproducing the issue, profiling with nsys and ncu, and consulting tooling references for targeted optimizations.

What is the best way to debug a slow CUDA kernel?

Debug a slow CUDA kernel using a standardized workflow that applies cuda-gdb and compute-sanitizer to diagnose issues. It helps developers reproduce problems, inspect kernel execution, and consult documentation to guide reproducible profiling and targeted optimizations.

When do I need to use Nsight Compute versus Nsight Systems for GPU profiling?

Use Nsight Systems for system-wide profiling to observe kernel timings and Nsight Compute for detailed kernel-level performance analysis. This workflow integrates both tools to provide a consolidated overview for diagnosing GPU kernels across a typical development cycle.

Can I inspect PTX ISA and binary information from compiled CUDA kernels?

Inspect compiled CUDA kernel binaries and PTX ISA using cuobjdump. This workflow includes binary inspection as part of a consolidated overview of kernel development, helping you understand low-level execution and diagnose performance issues.

Does this workflow support memory error detection in CUDA applications?

Memory error detection is supported through the compute-sanitizer tool. The workflow integrates compute-sanitizer alongside cuda-gdb and Nsight profilers to help developers identify, analyze, and resolve both performance bottlenecks and memory errors in GPU kernels.

Why should I standardize my CUDA optimization workflow?

Standardizing your CUDA optimization workflow ensures reproducible profiling and targeted optimizations across scattered tooling ecosystems. It provides a structured guide to debug, profile, and optimize GPU kernels consistently using standard tools like nsys, ncu, and cuda-gdb.