Why it matters

Nsight is the gold standard for CUDA perf. Skipping means guessing.

Advertisement

The architecture

Nsys captures timeline; ncu profiles kernels.

View gaps and stalls.

Identify slow kernels.

NVIDIA Nsight for LLMsNsight Systemstimeline captureNsight Computeper-kernel detailsReportgaps, stalls, hotspotsNsight is free from NVIDIA; use both tools together for best results
NVIDIA Nsight for LLMs.
Advertisement

How it works end to end

nsys profile python train.py.

ncu -k kernel_regex.

Open in Nsight GUI.