Jetson AGX Orin "System Throttled due to overcurrent" while running TensorRT

Hello,

I’ve read those threads :

-System throttled due to over-current?

-The error message "System throttling due to over-current" appears when running YOLOv4

-AGX Orin System throttled due to Over-current pytorch

So i know that the overcurrent issue is known by NVIDIA.
I have a Jetson AGX Orin Devkit and it runs fine, but when it is in the “max power” mode, running a GPU intensive model will trigger the warning. Running the same model in 50W will not trigger it, but the inference will be almost half the speed.
I tried different models, the bigger ones trigger the same warning, while the smaller one are working without issue.

I have 2 questions :

  1. When this warning is triggered, is the GPU frequency/power/etc… automatically limited (as the “throttled” seems to imply), and if not is there a risk to damage the hardware if this warning is triggered at every inference ?

  2. In the latest topic I found, AastaLLL from NVIDIA mentionned that a custom power mode could be created following this :
    NVIDIA Jetson Orin - Tuning Power - RidgeRun Developer Connection
    But what exactly should be limited to maximize GPU speed without triggering the overcurrent warning ?
    Is the overcurrent global and I would need to turn off/slow down some CPU cores, or is the overcurrent localized in the GPU and I need to limit its frequency (or else) ? (and then to which value ?)

Because given the number of parameters and possible combinations, benchmarking everything would be very long, and if every warning means putting a strain on the hardware, it wouldn’t be too good to do many experiments.

Thank you very much,
FV

Hi,

1. When the over-current happens, the clock will be reduced to protect the device.

2. Overcurrent is a system-wide issue.
You can reduce the CPU clock if most of the tasks are running on the GPU.

Thanks.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.