GPU Profiling (ncu)
Open the GPU performance counters on the host, and renters can run NVIDIA Nsight Compute (ncu) in their pods. It is one host driver flag; there is no portal switch. The node then gets a GPU Profiling badge and matches the renters' GPU Profiling (ncu) filter.
Why enable it
Few GPU clouds offer open counters, and kernel developers ask for them, most on B200 and B300.
8× H200 / B200 / B300 / GB300 nodes
Renters mostly rent these whole. Open the counters and list the node whole: the node keeps its idle pay and stays open to every whole-node renter.
A full 8× H200, B200 or B300 node earns the unrented incentive only with GPU Profiling, GPU splitting or an attested confidential VM.
The trade-off: whole-host rentals only
The counters are host-wide, so a profiling node is rented only as a whole host. GPU splitting is suspended while the flag is on.
Enable the counters
Change the flag only while the node has no active rental.
1. Set the driver option
echo 'options nvidia NVreg_RestrictProfilingToAdminUsers=0' | sudo tee /etc/modprobe.d/nvidia-profiling.conf
2. Rebuild and reboot
sudo update-initramfs -u
sudo reboot
3. Verify
grep RmProfilingAdminOnly /proc/driver/nvidia/params
It must print RmProfilingAdminOnly: 0. The node shows under the GPU Profiling (ncu) filter within an hour.
Disable the counters
On an idle node:
sudo rm /etc/modprobe.d/nvidia-profiling.conf
sudo update-initramfs -u
sudo reboot