Loud switching / coil whine noise on Jetson Orin Nano during inference

Robotics & Edge Computing Jetson & Embedded Systems Jetson Orin Nano
, , ,
1

I have been struggling with this issue for a while and was hoping for some insight. It seems this question is related to this post, which does not have a conclusive answer.

I am using Jetson Orin 8GB modules and a Nvidia Jetson Orin Nano Super carrier board for development. I run a pretty heavy pipelined gstreamer + CUDA + Deepstream python app on the Nano, which has 1 video stream as the input. There are two engines in this pipeline, each with a lot of output tensors, but only one of them processed with Deepstream. So obviously a heavy load for the Nano, but doable.

However, when I start testing this app, I hear a very loud coil whine sound. It is cyclic and happens in bursts. After some testing I determined it happens when using Deepstream, but you can most easily recreate it when running a larger model with trtexec (25W power mode is loud). The sound of the model changes with batch size, and model itself. I have seen this on multiple Jetson Orin Nanos and multiple carrier boards. I have included a spectrogram of the noise, as well as a recording of it below.

image
image1495×307 177 KB

Initially, I assumed it was coil whine, and that an inductor on the Nano Super carrier board was saturating due to high inrush current and the concurrent voltage drop. I tried multiple carrier board fixes. This included replacing the power inductors on the carrier board, which resulted in exactly the same sound, and attempting to simulate the load on the carrier board, which was produced no sound (though likely my e-load cannot recreate steep enough transients). This lead me to suspect the inductors on the Orin Nano itself.

For the test setup, I am using a Jetson Orin Nano 8GB, running r36.4.3 (JP6.2), and a custom heatsink. I have checked that there are no OCP events when this sound is happening on my power mode (and I am not using MAXN Super). The sounds is quieter at more conservative power modes, but still audible.

If anyone has any insight for what I can do mitigate this, that would be a huge help. I want the added performance of JP6.2, but right now the Nano is just way too loud.

5

For a bit of an update, We tried to load the GPU by running the 5_Simulations/nbody CUDA sample, which didn’t create the sound. If we ran the tensorrt engine at the same time the sound became audible again.
We tried running the tensorrt engine in debug mode, and every time we called"context.execute_async_v2(bindings=bindings, stream_handle=stream.handle)", you can hear single “click”.
So the fact that the cuda sample doesn’t produce the noise but the tensorrt engine does, points towards it only happening when utilizing the tensor cores of the Jetson Orin Nano.

6

System loading is use case dependent. The acoustic noise might be due to sudden variations in loading which may then fall in the audible range and you may need to characterize based on your use case. Acoustic noise damping measures/solutions may need to be applied at the system level.

7

Can you clarify what you mean by dampening at the system level? Are you suggesting adding external dampening solutions to the board? Also, could you give any indication of where the dampening solutions could be applied in that case? As mentioned previously, I was assuming that the tensor cores are allowed to pull very high inrush current, causing the primary power inductors to saturate and resonate. I have tried applying physical dampening to the inductors, and that did nothing to dampen the sound.

So then do you believe it the caps resonating? If so, which caps? What solutions would you recommend in this case, as most of the time (as far as I am aware), you need to add or change electronics to reduce this sort of noise.

If that is not what you mean by system level, please let me know what else I can do. So far, I have not been able to find any way to limit tensor core inrush current other than decreasing overall GPU frequency. Is there any other way I can limit usage more specifically? Thanks.