Nsight Compute 2026.2 - New Features

Added support for in-process profiling workflows that dynamically load the CUDA injection library. See the new CUDA injection API header extras/Api/nvComputeInj.h for usage details.
Added the --process-id command line option for --mode attach to attach to a specific process.
Added the --list-injection-path-64 and --list-injection-path-32 command line options to report the paths used to load the CUDA injection library.

Fixed an issue where profiling a green-context kernel could fail after a non-green-context kernel had already been profiled in the same process.

Fixed an issue that could cause ncu --mode attach to fail when range replay was followed by kernel replay in the same process.

Fixed a crash in the CUDA injection library when an application initialized CUDA before the injection library was loaded.

Added support for CUDA 13.3.
Added metrics for the size of SASS instructions.
Added profiling support for previously unsupported SKUs, with reduced metric support, including GPUs such as RTX 6000D BSE (GB202-891), RTX 5090DD (GB202-240), and RTX 5090D (GB202-250). For more details, refer to ERR_NVGPU.

Updated the overall layout to pin several tool windows by default.
Improved CUDA Tile support on the Source page.
Added a progress indicator for long-running source comparisons.
The Function Statistics tool window now shows per high-level source line data and the represented time range.
The Instruction Statistics section now shows warp-can't issue stall samples per HW warp ID slot.
Added a Device Information tool window to show information for local and remote systems from the Start Activity Dialog or the Connection menu.
Improved the Report Merge Tool dialog UI.
The CUDA Graph Viewer now shows the source of each graph. It also updates in the background without blocking the UI and provides progress information.
The Demangled Name column tooltip now always shows the full demangled name.

Improved the performance to first result when loading reports in the Python Report Interface.
Fixed several issues when merging reports.
Fixed UI crashes on Linux (aarch64 sbsa) platforms.
Fixed the achieved_fp32 formula in the SOL roofline rules.
Fixed an issue on the Source page that blank lines were highlighted in some cases for inlined functions.
Fixed failing SSH connections when the OpenSSH config contained unsupported escape sequences.

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

NVIDIA® Nsight™ Compute 2026.2 is available for download under the NVIDIA Registered Developer Program.