Introduction

Nsight Python is a Python kernel profiling interface that automates performance analysis across multiple kernel configurations using NVIDIA Nsight Tools.

With just a decorator and a context manager, you can benchmark kernel variants, collect detailed architectural metrics, prevent GPU throttling, and generate publication-quality visualizations—all from a single script. No boilerplate. No manual report parsing.

Built for researchers, engineers, and performance enthusiasts who want real metrics (not just wall-clock time), at scale, with minimal code overhead.

Key Features

• Works with any Python framework (PyTorch, Triton, etc.)

• Profiles multiple kernel configurations with one decorator

• Collects detailed GPU kernel metrics, not just timings

• Annotates kernel regions using a simple with nsight.annotate(…) context

• Prevents GPU throttling using real-time NVML temperature monitoring

• Automatically aggregates and visualizes performance results

• Outputs structured data (CSV, plots) for publication or analysis