Nsight Compute 2026.3 - New Features

Added support for CUDA 13.4.
Added support for the Rubin GPU architecture.
Added the Compute Triage Guide. It provides a top-down workflow for understanding compute workload performance bottlenecks and optimizations.
Added access to workload-aligned PM sampling timeline data from the Python Report Interface and the NvRules API.

Added a Tile IR view to the Source page for CUDA Tile workloads. This includes source to Tile IR, Tile IR to PTX, Tile IR to SASS correlation, CUDA Tile qualifier highlighting, and source file names for Tile IR entries.
Extended Source Comparison support to include Tile IR source code comparison.
Added an option on the Source page to filter the SASS view to the instructions generated by the selected source line.
Improved the CUDA Graph Viewer with a legend, clickable profiled nodes, persisted pan and zoom state, improved long-name handling, and better visualization of nested graph structures.
Added recent remote files and folders to the remote file dialog.
Improved PM sampling timelines with metrics in group rows, merged timelines from matching section bodies, and more useful interactive tooltips.
The Details page now merges matching expansion section bodies into parent sections.
Nsight Copilot can now check for updates, download and verify update artifacts, install updated packages, and manage update behavior in the Options dialog.
The Nsight Copilot UI improved installation, references, saved chat files, long inputs, and authentication hints.
Added a CUDA tab to the Device Information tool window showing CUDA device attributes such as compute capability, memory, thread and block limits, registers and shared memory limits, compute feature support, and other device properties.

Added Tile IR source output and source to Tile IR, Tile IR to PTX, Tile IR to SASS correlation to the source page console output.
Added --list-injection-path-32 and --list-injection-path-64 to print the default injection library directories.

Isolated the embedded Python interpreter from the system Python environment, unless NV_NSIGHT_PYTHON_ISOLATED=0 is set.
Fixed an issue where profiling through a shell wrapper could be significantly slower than profiling the wrapped binary directly.
Fixed an issue with profiling CUDA Green Contexts after profiling on the primary context.
Fixed duplicate source files in the Source page dropdown for CUDA Tile range and graph profiling reports.
Fixed an issue where the CUDA Graph Viewer could keep showing previous graph content after profiling a non-graph workload.
Fixed that metric tool window instance bars intermittently didn’t match the actual values.
Fixed incorrect units for user metric expressions in section files.
Fixed an issue where rule results could be missing when merging reports by section set.
Fixed the Report Merge Tool file list to prevent editing currently open report names.

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

NVIDIA® Nsight™ Compute 2026.3 is available for download under the NVIDIA Registered Developer Program.