Performance on AGX Orin

I was looking to use ZED X cameras paired with a AGX Orin compute for my application.

I wanted to know how much resources would the ZED SDK take up in terms of GPU%, CPU% and Memory by each camera for computing the depth using the NEURAL mode or ULTRA mode with HD1080 resolution. Besides that, I also wanted to know what kind of FPS I can expect with these modes.

I want to know this so that I can calculate if there would be sufficient compute available to run other models and algorithms on the Orin while the cameras are running.

I saw some information here, which indicated for NEURAL mode the GPU usage is around 80%, which would make it difficult to run any other models with parallel or even run a 3rd ZED camera with NEURAL mode.


You can run more than 2 cameras with the current NEURAL mode. With 2 cameras, if you don’t have anything to heavy on the CPU and GPU, you reach 30 FPS. If you want to put algorithms beside, it will reduce the ZED FPS. You can also use the grab_compute_capping_fps (InitParameters Class Reference | API Reference | Stereolabs) parameter to forcefully reduce the camera FPS, and give more space to external programs.

Thanks for the information!
I know there are a lot of possible configurations for the camera and depth parameters, but do you have any benchmarks you maintain internally about how much compute resources the SDK takes for depth computation for certain configurations?

Unfortunately, we cannot run these kind of benchmark for all configurations, there are too many.
I can also tell you that you can reach 15FPS with 4 cameras, it’s pretty linear. The memory should not be a problem. As for the CPU and GPUs, you can reduce their usage by capping the FPS.

Thanks for the information @alassagne!

1 Like