Question regarding Jetson Orin Nano Super system throttling due to overcurrent/undercurrent

Robotics & Edge Computing Jetson Systems Jetson Orin Nano
1

I am currently running an inference file (.py) on a Jetson board after converting it to a .engine file.
The power settings are: power mode: 15W, and I am using the sudo jetson_clocks configuration.

  1. When I run the inference file (.py) under the above settings, no throttling occurs. However, when the power settings are set to Power Mode: 25W and I use sudo jetson_clocks, throttling occurs.
  • It seems the cause of the throttling is that during inference, the system detects a clock speed exceeding the configured limit, causing a peak in overcurrent.
  1. My question is: Is it okay to ignore this throttling and continue using the board?
  • Will the board break if I ignore it? (When set to power mode: 25W and sudo jetson_clocks, the throttling count is around 200–300; when set to power mode: MANX SUPER and sudo jetson_clocks, the throttling count reaches 800 per execution)
    In your opinion, what is the threshold for a dangerous throttling count?
  1. If I can’t ignore it, do I need to set a power limit or create a custom power management scheme? Are there any other options?
4

Hi suseong326,

Are you using the devkit or custom board for Orin Nano?
What’s the Jetpack version in use?

You may get performance decreased if you ignore it.
Please refer to Jetson Orin Nano Series, Jetson Orin NX Series and Jetson AGX Orin Series — NVIDIA Jetson Linux Developer Guide for OC event.

5

I used Jetson orin nano super 8GB Developer Kit, jetpack ver :6.1.2

6

Okay, you can simply use the SDK manager updating to the latest Jetpack 6.2.2(r36.5) for the devkit.

Please let us know if you have further question about the OC event and the throttling.

7

Power Mode: 15W; I am running the inference program using the sudo jetson_clocks configuration.

  1. With the current settings, no throttling occurs. However, I discovered that throttling occurs (about 300 times per inference) when using Power Mode: 25W. In this case, is using the current configuration better for inference performance? (I found that when throttling occurs after setting it to 25W, performance drops by 50%)
  2. Alternatively, is it okay to use PowerEstimator to increase the GPU and CPU clock speeds so that throttling (oc3) does not occur at 15W?
8

Yes, we recommend utilizing the PowerEstimator to generate a custom power mode configuration tailored to your use case. This will prevent OC throttling and enable you to achieve optimal performance.