Profile CPU-GPU timelines, API calls, OS events, and NVTX ranges.
npx skills add https://github.com/NVIDIA/skills --skill perf-nsight-systems
Please help me install this Agent Skill. Skill: perf-nsight-systems Source: https://github.com/NVIDIA/skills/tree/main/skills/TensorRT-LLM/perf-nsight-systems Command: npx skills add https://github.com/NVIDIA/skills --skill perf-nsight-systems
Nsight Systems profiles system-level performance to reveal how CPU and GPU work together over time, enabling you to pinpoint scheduling bottlenecks, overlap gaps, and data movement issues.
Profile a representative training run with Nsight Systems, generate an .nsys-rep, and begin analysis with nsys stats/analyze/recipe.
High-intent search queries and answers about installing and using this skill.
System-level timeline profiling reveals how CPU and GPU work together over time, capturing API calls and OS events to pinpoint scheduling bottlenecks and data movement issues during deep learning training.
You generate timeline traces with NVTX annotations by profiling a representative run to create an .nsys-rep file, which correlates code with GPU work through NVTX ranges and PyTorch autograd integration.
Yes, Nsight Systems supports distributed training and inference workflows, utilizing PyTorch autograd integration to correlate code with GPU work and capture CPU and GPU activity timelines.
The best way to analyze GPU performance after generating an .nsys-rep file is to use nsys stats, analyze, and recipe commands for post-run analysis, enabling you to export detailed reports.
System-level GPU profiling focuses on how CPU and GPU work together over time to reveal scheduling bottlenecks and overlap gaps, rather than focusing solely on individual kernel execution metrics.
Use system-level timeline profiling when you need to identify CPU-GPU scheduling bottlenecks, overlap gaps, and data movement issues in DL training, distributed training, or inference workflows.
yo-stevenInstrument PyTorch training loops and CUDA kernels to measure latency and throughput.
yo-stevenProfile CPU, GPU, and communication timelines from .nsys-rep captures.
jstzwjQuantize weights and optimizer states for PyTorch model training.
wannasaynoneProfile CPU, GPU, memory, and load times to identify game performance bottlenecks.
