NVIDIA released Nsight Python, an open-source Python interface designed to profile GPU kernels directly within environments like Jupyter and Colab. Instead of relying solely on command-line tools and separate reports, developers can now mark target code sections with nsight.annotate and use nsight.analyze.kernel to generate a pandas DataFrame containing hardware metrics per run and configuration.
The library attaches Nsight Compute directly to a running Python process, streamlining the collection of hardware metrics, cache performance, clock control, and matrix size sweeps in fp32 and fp16. The recent update introduces features such as an experimental CUPTI backend, explicit tool management, and custom information collectors, allowing developers to easily distinguish between compute-bound and memory-bound workloads.
This tool improves the efficiency of GPU kernel optimization by enabling direct performance analysis inside standard Python workflows.