Updates in 2020.2

General

  • Added support for the NVIDIA Ampere GPUs with compute capability 8.6 and CUDA toolkit 11.1.

  • Added support for application replay to collect metric results across multiple application runs, instead of replaying individual kernels.

  • Added new launch__device_id metric.

  • Added support for NVLink (nvl*) metrics for GPUs with compute capabilities 7.0, 7.5 and 8.0

  • Added documentation for memory charts and tables in the Profiling Guide.

NVIDIA Nsight Compute

  • Updated menu and toolbar layout.

  • Added support for zoom and pan on roofline charts.

  • The Resources tool window shows the current CUDA stream attributes.

  • The memory chart shows a heatmap for link and port utilization.

  • The hot-spot tables in the Source Counters section now show values as percentages, too.

  • On-demand resolve of remote CUDA-C source is now available for MacOS hosts.

  • Metric columns in the Summary and Raw pages are now sortable.

  • Added a new option to set the number of recent API calls shown in the API Stream tool window.

NVIDIA Nsight Compute CLI

  • CLI output now shows NVTX payload information.

  • CSV output now shows NVTX states.

  • Added a new --replay-mode option to select the mechanism used for replaying a kernel launch multiple times.

  • Added a new --kill option to terminate the application once all requested kernels were profiled.

  • Added a new --log-file option to decide the output stream for printing tool output.

  • Added a new --check-exit-code option to decide if the child application exit code should be checked.

Resolved Issues

  • The profiling progress dialog is not dismissed automatically anymore after an error.

  • The inter-process lock is now automatically given write permissions for all users.

  • All project extensions are enabled in the default dialog filter.

  • Fixed handling of targets using tcsh during remote profiling.

  • Fixed handling of quoted application arguments on Windows.