How to verify correct GPU usage for Sionna RT?

[studycreate11]

Hello,

I have a question about how to properly use the GPU with Sionna RT.

I'm currently running the Sionna RT tutorials on my Linux system.

When I execute the code, I've been monitoring my GPU with nvidia-smi, and I see the GPU-Util metric rise to about 15-25%.

Is it possible that the GPU utilization is low simply because the code's task is too simple for the GPU?

Any clarification would be greatly appreciated.

Thank you!

[jhoydis]

Hi @studycreate11,

I do not know how many transmitters you in the scene, but 10M samples per TX is not a lot for radio map computation. Also your cell size of 100mx100m is quite coarse. I am not surprised that this does not max out the GPU utilization. You could also use nvtop to see a more detailed visualization of the GPU utilization over time (You can install it like this: sudo apt install nvtop).

[studycreate11]

Thanks for your help!

[studycreate11]

Sorry for asking again
I checked with nvtop and saw that my CPU usage goes up to 1800% when loading a scene. I think this means Sionna RT is running on the CPU. Are there any settings I need to configure first to make it run on the GPU? I've looked into how to run it on the GPU, but it's not working, so I'm asking for help. My current device is an NVIDIA GeForce RTX 3090.

[merlinND]

Hello @studycreate11,

Before rm_solver = RadioMapSolver(), try running the following code:

import mitsuba as mi
print(mi.variant())

If the variant starts with cuda_, then the GPU should be used. You can try changing the parameters of the simulation in order to increase the amount of GPU work vs the CPU overhead of loading and setting up the scene.

[studycreate11]

The output was cuda_ad_mono_polarized.

So, I guess the GPU utilization was low because the code's task was too simple.