Nsight Compute 2023.2 - New Features

Fixed possible crash when profiling CUDA graphs on multiple GPUs.
Fixed the detection mechanism of the C2C interface, i.e. metric c2clink__present. The fix requires the display driver shipping with this release or any newer driver.

Fixed a crash during application replay while having the temporary directory located on a network file system (NFS).
Improved detection mechanism for C2C interface. Added caching of the detected configuration to reduce overhead.

Extended the rules system to show estimates of the potential speedup that can be achieved by addressing the corresponding performance bottleneck. These speedups allow prioritizing applicable rules and help focusing first on optimization strategies with the highest potential performance gain.
Added support for rules to highlight individual source lines. Lines with global/local memory access with high excessive sector counts and shared accesses with many bank conflicts are automatically detected and highlighted.
Added the ability to query metric attributes in NvRules API.
Added support for creating instanced metrics through the NvRules API.
Added sharedBankConflicts sample CUDA application and document to show how to use NVIDIA Nsight Compute to optimize shared memory access patterns.
For Orin+ mobile chips on the Linux aarch64 platform, added metrics (mcc__*) support for memory controller channel (MC Channel) unit which connects to the DRAM.

Added hyperlinks to the SASS View of the Source Page for instructions that reference others by address or offset. This enables to quickly jump to the target instruction of a branch.
Improved the search bar in the Metric Details tool window. The search string now matches any part of the metric names, and the matching results are shown in a sorted order.
Added a visual indication of scale of the metric value changes when the baselines are used. The background bars in the table cells of the Details Page allow to quickly identify which metrics values increased or decreased the most. The color scheme can be configured in the Baselines tool window.
Added support for rules toggle button on the Summary Page. Allows to hide the bottom pane with the rules output for the selected kernel launch.
Added support for allowing users to configure properties on Summary Page using Metrics/Properties profile option.
Added percentage bars on Summary Page.

Added support for tracking child processes launched with posix_spawn(p) when using --target-processes all.
Added support for tracking child processes launched with ystem() on Windows and Linux (aarch64, x86_64) when using --target-processes all.

Fixed table alignment in the output of the NVIDIA Nsight Compute CLI on Windows when printing Unicode characters.
Fixed view corruption in the Source Page after switching from the collapsed view to the expanded view.
Fixed missing tooltip descriptions for some SASS instructions.
Fixed potential crash when copying from the Resources tool window using CTRL+C.
Fixed possible crash wh

For a complete overview of all NVIDIA Nsight™ Compute features and access to resources, please visit the main Nsight™ Compute page.

NVIDIA® Nsight™ Compute 2023.2 is available for download under the NVIDIA Registered Developer Program.

References