Nsight Compute 2021.2 - New Features
Updates in 2021.2.2
- Support for CUDA Toolkit 11.4 Update 2 was added.
- Enhanced support for NVIDIA virtual GPUs (vGPUs) in the upcoming GRID release.
Updates in 2021.2.1
- Support for CUDA Toolkit 11.4 Update 1 was added.
- Reduced the memory overhead when loading reports in the Python Report Interface.
- Fixed that links in the Memory Allocations Resource view were not working correctly.
- Fixed that NVTX state might not be correctly reset between interactive profiling activities.
- Fixed that the UI could crash when opening baselines from different GPU architectures.
Updates in 2021.2.0
- Added support for OptiX version 7.3.
- Added support for profiling on NVIDIA virtual GPUs (vGPUs) on an upcoming GRID/vGPU release.
- Added a new Python-based report interface for interacting with report files from Python scripts.
- Added a new rule to warn users when sampling metrics were selected, but no sampling data was collected.
- Renamed SOL to Throughput in the Speed of Light section.
- Renamed several
memory_*metrics used on the Source page, to better reflect the measured value. See the Source page documentation for more details.
NVIDIA Nsight Compute
- Added support for opening cubin files in a Standalone Source Viewer without profiling the application.
- Moved the output of all rules so that it is visible even if a section's body is collapsed. Visibility of the rules' output can be toggled by a new button in the report header.
- The profiler report header now shows the report name for each baseline when ambiguous.
- Rules can define Focused Metrics that were most important for triggering their result output. Metrics are provided per result message which additional information, such as the underlying conditions and thresholds.
- Memory tables show tooltips for cells with derived metric calculations.
- Added a knowledge base service to show more comprehensive background information on metric names and descriptions in their tooltips.
- Following a link in the Source Counters hot spot tables automatically selects the corresponding metric in the Source page.
- Added new columns for visualizing register dependencies in the SASS view of the Source page.
- Functions in the SASS view are now sorted by name.
- Added support for OptiX 7.x resource tracking in the interactive profile activity. The Resources tool window will show information on instantiated
optixDeviceContexts, optixModules, optixProgramGroups, optixPipelinesand
- Added support for new CUDA graph memory allocation APIs.
- Improved consistency between command line parameters and the Next Trigger filter in the API Stream window for handling of regex inputs. The Next Trigger filter now considers kernel/API name as a regular expression only if string has
- Added ability to select font settings in the options dialog.
- Added ability to configure the metrics shown on the summary page via the options dialog.
- The selected heatmap color scale now also applies to the Memory chart.
- The ncu-ui script now checks for missing library dependencies, such as OpenGL or Qt.
NVIDIA Nsight Compute CLI
- Added environment variable
NV_COMPUTE_PROFILER_DISABLE_STOCK_FILE_DEPLOYMENT=1to skip deployment of section and rule files.
- Fixed a performance issue in the NVIDIA Nsight Compute CLI when using
--page raw --csv --units auto.
- Fixed that the SSH passphase key is no longer persisted in the project file.
- Fixed state of restore button in connection dialog. The button now supports restoring the default settings, if current setting differ from the default.
- Fixed that the complete GPU name can be shown in the NVLINK topology diagram on MacOS.
- Fixed that collapsing the Source view reset the selected metrics.
- Fixed that correlated lines could differ between filtered and unfiltered views of the executed functions.
- Fixed that two application icons were shown in the MacOS dock.
- Improved HiDPI awareness.
For a complete overview of all NVIDIA Nsight™ Compute features and access to resources, please visit the main Nsight™ Compute page.
NVIDIA® Nsight™ Compute 2021.2 is available for download under the NVIDIA Registered Developer Program.
Download 2021.2.2 Download 2021.2.1 Download 2021.2.0 Documentation