Nsight Python 1.1: Explicit Tool Management and CUPTI Backend Enable GPU Profiling in Jupyter and Colab

NVIDIA has introduced new capabilities in its upcoming Nsight Python 1.1 release that allow developers to profile GPU kernels directly from Python code without leaving their development environment. According to MarkTechPost, this open-source interface enables teams to analyze hardware metrics, evaluate cache controls, and measure kernel fusion by marking specific code regions for inspection. The update aims to streamline the process of gathering performance data from complex GPU operations.

How it unfolded

• NVIDIA released version 1.0 of Nsight Python, which attached the Nsight Compute profiler to active Python processes for use in Jupyter and Colab.
• The upcoming 1.1 pre-release introduces an experimental CUPTI backend, explicit tool management, and custom information collectors for deeper analysis.
• MarkTechPost documented a tutorial using these tools on a free Colab T4 instance to profile matrix multiplication and measure hardware counters.

Who is affected

Developers working with NVIDIA GPUs who need to optimize kernel performance are the primary audience for this tooling update. According to MarkTechPost, the library is particularly useful for those running code in cloud environments like Google Colab, where default profiler versions may be outdated. Users must manage their own installation of Nsight Compute, ensuring they have version 2026.2.1 or later available on their system path for the Python interface to function correctly.

What happens next

Teams adopting the 1.1 pre-release will need to handle package dependencies carefully, especially when using environments with pinned CUDA libraries. MarkTechPost notes that the cupti-python package requires a newer CUPTI library than standard PyTorch installations in Colab, necessitating separate directory installations. Developers must also ensure the nsight module is imported before any CUDA workloads start, as the injection library only attaches to GPU activity initiated after the import.

The background in plain language

Profiling GPU kernels involves measuring how efficiently hardware executes specific computational tasks. Developers use these metrics to identify bottlenecks, such as memory bandwidth limits or underutilized compute units. By attaching profilers directly to Python processes, teams can capture real-time data without stopping their workflow. This approach allows for iterative testing, where developers can adjust code parameters and immediately see how changes impact hardware performance, leading to more optimized and efficient applications.

What to do and how to stay safe: NVIDIA

• Verify that your development environment has the required version of Nsight Compute installed and accessible via the system path.
• Ensure the nsight module is imported at the very beginning of your script, before any CUDA libraries or GPU operations are initialized.
• Check for compatibility issues between your CUDA toolkit version and the profiling libraries, particularly in managed cloud environments.
• Review the documentation for the experimental CUPTI backend to understand its current limitations and data collection capabilities.

Step-by-step guide: Mobile Device Security for Small Business: Control Without Cost

General security guidance from the Circuit Observer newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

What is the minimum version of Nsight Compute required for Nsight Python?

Nsight Python requires Nsight Compute command-line tool version 2026.2.1 or later to function correctly.

Can Nsight Python be used in Jupyter or Google Colab?

Yes, since version 1.0, the library attaches to running Python processes, making it usable in Jupyter and Colab environments.

Why must the nsight module be imported before CUDA work?

Importing nsight loads the injection library into the Python process, which can only attach to CUDA work started after that specific import.