Unlocking Performance Insights with NVIDIA Nsight Compute

Summary

NVIDIA Nsight Compute is a powerful tool designed to help developers optimize and debug CUDA applications. This article explores the latest features and improvements in Nsight Compute, focusing on enhanced performance visualization and guidance. We’ll delve into the new features, such as improved tooltips, enhanced source syntax highlighting, and the addition of Python Call Stacks, to understand how these updates can streamline the development process.

Introduction

Developing high-performance CUDA applications requires detailed insights into how your code interacts with the GPU. NVIDIA Nsight Compute is an interactive kernel profiler that provides these insights, helping developers identify and fix performance issues. With its latest updates, Nsight Compute offers even more tools to visualize and optimize application performance.

Enhanced Performance Visualization

One of the key updates in Nsight Compute is the improved performance visualization. This includes:

  • PM Sampling Timelines: These now show sampled GPU workload activities, providing a clearer picture of how your application utilizes the GPU.
  • Improved Tooltips: Enhanced tooltips in the memory chart offer more detailed information, especially when metrics are missing.
  • Redesigned Report Header: The new report header layout makes it easier to access all report pages and perform actions like adding a baseline.

Enhanced Source Analysis

Nsight Compute also includes several updates to enhance source analysis:

  • Source Page and Source Comparison: These have been redesigned to allow more vertical space and include features like linked dropdowns for easier navigation.
  • Inline Table Support: Added to the Source Comparison document, this feature allows for more detailed analysis.
  • Rich Syntax Highlighting: Support for Python and Fortran source syntax highlighting has been added, along with enhanced CUDA-C and PTX syntax highlighting.

Python Call Stacks

A significant addition is the support for collecting Python Call Stacks alongside native ones. This feature helps developers better understand the context of a workload in Python applications, making it easier to identify performance bottlenecks.

Acceleration Structure Viewer

The Acceleration Structure Viewer has been enhanced to compute ray-geometry intersection and traversal timing heatmaps. This tool is invaluable for inspecting acceleration structures used in ray-tracing pipelines, helping developers identify inefficiencies and errors.

Custom Metric Descriptions

Nsight Compute now allows for specifying custom metric descriptions in section files. This feature provides more flexibility in how metrics are displayed and interpreted.

Improved Handling of Short Workloads

The handling of short workloads during PM sampling has been improved, ensuring that even brief workloads are accurately captured and analyzed.

Fixed Issues

Several issues have been resolved, including problems with report saving, PM sampling reporting incomplete data, and issues with profiling multi-ctx applications on vGPU.

Key Features Overview

Feature Description
PM Sampling Timelines Shows sampled GPU workload activities.
Improved Tooltips Enhanced tooltips in the memory chart.
Redesigned Report Header Easier access to all report pages.
Source Page and Source Comparison Redesigned for more vertical space.
Inline Table Support Added to Source Comparison document.
Rich Syntax Highlighting Support for Python, Fortran, CUDA-C, and PTX.
Python Call Stacks Collects Python Call Stacks alongside native ones.
Acceleration Structure Viewer Computes ray-geometry intersection and traversal timing heatmaps.
Custom Metric Descriptions Allows specifying custom metric descriptions in section files.
Improved Handling of Short Workloads Accurately captures and analyzes brief workloads.

Additional Resources

For a complete overview of all NVIDIA Nsight Compute features and access to resources, please visit the main Nsight Compute page. NVIDIA Nsight Compute is available for download under the NVIDIA Registered Developer Program.

Conclusion

NVIDIA Nsight Compute continues to evolve as a powerful tool for optimizing and debugging CUDA applications. The latest updates offer enhanced performance visualization, improved source analysis, and new features like Python Call Stacks and custom metric descriptions. By leveraging these tools, developers can gain deeper insights into their applications’ performance and make targeted optimizations to achieve better results.