NVIDIA® Nsight™ Compute 2021.1.0 is now available for download in the NVIDIA Registered Developer Program.
Version 2021.1.0 supports the CUDA Toolkit 11.3, Optix 7 API, and NVIDIA’s latest Ampere architecture GPUs. The Resources view supports new CUDA 11.3 memory pool allocations, CUDA graph user objects , and stream captured CUDA graph nodes . The NvRules API adds support for NVTX states, while the NVLINK section now includes a topology chart and link properties. Additionally, there are numerous other workflow enhancements, UI improvements, performance upgrades, and bug fixes.
- Added support for the CUDA toolkit 11.3, including
Nsight Compute Resource Tracking for new CUDA 11.3’s features:
- CUDA Graph’s User Objects
- Stream Ordered CUDA Memory Allocator’s memory allocations from Memory Pools
- Stream Captured CUDA Graph Nodes
- Added support for the OptiX 7 API.
GpuArchenumeration values used for filtering in section files were renamed from architecture names to compute capabilities.
- NVTX states can now be accessed via the NvRules API.
- Added a rule for the Occupancy section.
- Added support for new CUDA asynchronous allocator attributes in the Memory Pools resources view.
- Added a topology chart and link properties table in the NVLink section.
- The selected metric column is scrolled into view on the Source page when a new metric is selected.
- Users can choose the Source heatmap color scale in the Options dialog.
- Added file-based application replay as the new default application replay mode. File-based replay uses a temporary file for keeping replay data, instead of allocating them in memory. This keeps the required memory footprint close to constant, independent of the number of profiled kernels. Users can switch between buffer modes using the
- CLI output now shows NVTX color and message information.
--kernel-regex-baseoptions are deprecated and replaced by
- All options which support regex need to provide
regex: as a prefix before an argument to match per the regex, e.g
- Fixed that baselines were not updated properly on the Comments page.
- Fixed that NVTX ranges named using their payloads can be used in NVTX filtering expressions.
- Fixed crashes in MacOSX hosts when terminating the target application.
- The NVLINK(
nvl*) metrics are now added back.
For a complete overview of all NVIDIA Nsight™ Compute features and access to resources, please visit the main Nsight™ Compute page.