Issue fully described here:
If you are reporting a bug or error, consider submitting a Support Bundle to aiworkbench-ea@nvidia.com. This will help us solve your issue more quickly.
Please describe your issue or request: (tick the boxes after creating this topic):
Please tick the appropriate box to help us categorize your post
Bug or Error
Feature Request
Documentation Issue
Other
Confirmed — this is the CAP_SYS_PTRACE issue. nsys needs to attach to the CUDA process and trace kernel launches. Without CAP_SYS_PTRACE, the profiler falls back to CPU/OS sampling and the CUDA lane is empty.
Two ways to fix it inside AI Workbench:
1. If you launch the container manually, add --cap-add=SYS_PTRACE to the docker run command (or cap_add: [SYS_PTRACE] in compose).
2. If AI Workbench manages the container, add the capability to the project’s environment spec. In Workbench, this is typically done via the Dockerfile or container config depending on the project template version.
Also verify you are using the Nsight Systems build that matches the container’s CUDA toolkit, not the host’s. GB10 is aarch64 + CUDA 13; mixing x86_64 or CUDA 12 Nsight binaries will give you a report file but no CUDA data.
One more subtle thing: on UMA systems like GB10, nsys sometimes needs --trace=cuda,nvtx,osrt rather than just --trace=cuda to correlate GPU work with the CPU launch side correctly.