Can it help with specific NVIDIA architectures?

Yes, it provides optimization strategies tailored to specific architectures like Ampere, Hopper, and Ada Lovelace to leverage generation-specific hardware features.

Does it require external profiling tools to work?

While it leverages knowledge from tools like Nsight Copilot, it performs static analysis directly on your code to identify performance patterns before you even run a hardware profiler.

How does this skill help with GPU memory issues?

It analyzes your code for uncoalesced memory access patterns and shared memory bank conflicts, providing specific refactoring suggestions to improve bandwidth utilization.

What file types does it support?

It primarily focuses on CUDA source files (.cu) and headers (.cuh), but can also provide guidance on general GPU compute patterns in C++ or Python-based frameworks.

CUDA Performance Optimizer

Name: CUDA Performance Optimizer
Author: keith-mvs

bykeith-mvs

•

데이터 과학 및 ML

Optimizes CUDA kernels and GPU code to maximize hardware utilization and execution speed.

소개

The CUDA Optimization skill provides specialized expertise for auditing and enhancing GPU-accelerated code within the Claude Code environment. It automatically analyzes CUDA kernels to detect uncoalesced memory accesses, warp divergence, and suboptimal thread configurations while suggesting architecture-specific improvements for modern NVIDIA generations like Ampere, Hopper, and Ada Lovelace. By integrating deep profiling knowledge, it helps developers transition from functional GPU code to highly efficient, production-grade parallel algorithms through actionable code transformations and bottleneck identification.

주요 기능

1 GitHub stars
Memory access pattern analysis for coalescing and bank conflicts
Strategic shared, constant, and texture memory implementation
Architecture-specific tuning for NVIDIA GPU generations
Warp efficiency and branch divergence detection
Thread block and occupancy optimization for maximum throughput

사용 사례

Identifying memory bottlenecks in large-scale parallel algorithms
Refactoring slow CUDA kernels to improve execution speed
Optimizing GPU resource utilization for high-performance computing tasks

소개

주요 기능

1 GitHub stars
Memory access pattern analysis for coalescing and bank conflicts
Strategic shared, constant, and texture memory implementation
Architecture-specific tuning for NVIDIA GPU generations
Warp efficiency and branch divergence detection
Thread block and occupancy optimization for maximum throughput

사용 사례

Identifying memory bottlenecks in large-scale parallel algorithms
Refactoring slow CUDA kernels to improve execution speed
Optimizing GPU resource utilization for high-performance computing tasks