Nsight Compute 2026.4 - New Features

Updates in 2026.4.0

General

Added support for CUDA 13.5.
Added CUDA Tile kernel profile support on Windows on Arm.
Added access to Tile IR source correlation from the Python Report Interface and the NvRules API.
Added access to CUDA graph topology, node parameters, dependencies, and the graph node associated with each profiled result from the Python Report Interface.
Added support for profiling and replaying NVLink utilization split CUDA graph kernel nodes on Rubin GPUs.
Added limited support for C2C metrics for GB10B.
Added application replay support when profiling with the shmem communicator.
Added the launch__count metric to identify results without kernel launches. The UI and CLI now report a clear warning when no launches were collected.
Renamed the Rubin UTCMMA PM Sampling section files to High/Low Precision, since they apply to both training and inference.

NVIDIA Nsight Compute

Nsight Copilot can now explain Source page markers using the /explain Marker command.
Added search functionality in Nsight Copilot. The active chat can be searched by case-sensitive, whole-word, or regular expression matches.
Nsight Copilot responses now provide direct links to report metrics and source lines.
Added a Timeline page that consolidates timelines previously shown on the Details page. The page includes analysis views for function statistics and statistics for timeline rows over a selected time range.
Added support for visualizing one-dimensional and three-dimensional batched memcpy nodes in the CUDA Graph Viewer, including their parameters and dependencies.
Added an RT Live State view to the Source page for OptiX workloads. The view groups live state values by ray tracing callsite or source definition, reports live state byte and value counts, and links to associated source lines and SASS save/reload instructions when available.
Added support for importing in-memory source passed to NVRTC APIs for source correlation.
Added support for splitting Source page analysis views into two side-by-side panes.
The Source page SASS filter tooltip now shows the number of filtered lines.
The Source page now automatically selects a navigation metric when it is the only metric enabled from the column chooser.
The Source page vertical header rows now use interactive tooltips.
For Green Context results, added the locality-domain ID and SM resource flags to the report header, the Session page’s Green Context Resources table, and the Green Context Resources view in the Resources tool window.
Added information on Zero Bandwidth Clear (ZBC) compression to the memory chart L2 Compression node tooltip.
Restructured the Resources tool window. A categorized navigator with one column per domain and a search box replaces the single-select dropdown, and each resource opens in its own dockable window so several can be inspected at once. The default dock position for single and bulk opens is configurable from the navigator’s settings.
The CUDA context table in the Resources tool window now shows the current and minimum per-block memory sizes.
Graph memory-allocation nodes now report all CUDA memory location types, including host, host-NUMA, and device-locality-domain, as well as the managed allocation type. The Resources tool window, the CUDA Graph Viewer tooltip, and the Python Report Interface previously showed ??? for these values.
Added Source View Profiles to the Source page column chooser.
Improved the remote SSH connection creation and authentication workflow. Passwords and key passphrases are prompted for when needed and can be saved after successful authentication.
The Device Information System tab now groups information into sections and shows profiling permissions and CUDA driver compatibility.
Added user-configurable report page visibility through the View menu and page tab context menus. Visibility preferences are persisted, and navigation automatically reveals a hidden destination page.

NVIDIA Nsight Compute CLI

Invalid values for command-line options with a fixed set of choices now report the supported values.

Resolved Issues

Fixed several Nsight Copilot issues.
Fixed CUDA device graph launch profiling failures and hangs, including failures on Windows and Blackwell GPUs.
Fixed a crash when profiling node-level data for device-side graph launches in relocatable device code applications. Unsupported launches are now skipped with a warning with driver 620 or later.
Fixed missing or incorrect PM sampling metrics and rule results after merging reports.
Fixed hangs during collection when profiling PM sampling metrics with very large buffer sizes.
Reduced hangs of the UI when loading large report files.
Fixed the UI becoming unresponsive when closing a report, or when switching the focused result or source view, while background source correlation was still running.
Fixed the Metric Details window not opening when clicking a metric link in a metric instances table tooltip.
Fixed custom tooltips covering the control or table content being inspected, and being dismissed before the cursor could reach them.
Fixed a crash when the API Stream displayed a graph memory-allocation node carrying nested pointer arrays, such as access descriptors or external-semaphore handles.
Fixed duplicate file names in the Tile IR source dropdown for range replay reports.
Fixed stale source correlation highlights when selecting a line without source correlation.
Fixed incorrect estimated speedup values in the Summary page and rule results for some kernels.
Fixed negative SM occupancy and warp stall values when profiling some workloads on GR100.
Fixed issues when profiling kernels launched with cudaDevAttrCooperativeLaunch and cudaDevAttrClusterLaunch.
Fixed several input validation, resource utilization, and zero shared-memory handling issues in the Occupancy Calculator Python Interface.
Fixed memory leaks in the Python Report Interface when accessing IAction objects.
Added a warning to the High Pipe Utilization rule on GB10B when the XU utilization metric might overstate utilization for kernels using MUFU.EX2 instructions.
Improved error reporting when process attachment fails because of an incompatible LD_PRELOAD configuration.
Fixed a potential crash when launching a profiling session for an application that does not exist on a Linux or WSL remote target.

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

NVIDIA® Nsight™ Compute 2026.4 is available for download under the NVIDIA Registered Developer Program.

Download 2026.4 Documentation