Hey,
I’m trying to profile a Pytorch script that utilizes CUDA capabilities to observe GPU workload percentage of the Jetson AGX Orin platform. But when I execute the command below to see the supported devices for --gpu-metrics-device option, it says that Orin is not supported.
nsys profile --trace=cuda --cudabacktrace memory --cuda-memory-usage true --gpu-metrics-device help python3 stream-util.py
Possible --gpu-metrics-device values are:
0: Orin (not supported)
all: Select all supported GPUs
none: Disable GPU Metrics [Default]
Some GPUs are not supported. See the user guide: https://docs.nvidia.com/nsight-systems/UserGuide/index.html#gms-introduction
The user guide it’s referring to does not contain any information about this situation. What else I can do to observe GPU workload percentage for a Pytorch script?
I’m aware of the tegrastats software but I believe it runs for the entire system and there is no way to specify a script for measurement, right?