GPU CUDA loading via ssh

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello,

Me and my team work with Nvidia Jetson Nano microcomputer. We designed baseboard especially for our usage. There is no hdmi/displayport or regular usb ports. We need to place it into aluminium enclosure along with rest of our electronics and there is no space for heatsilk. We use our enclosure for heat distribution and it works well for now. In the future we have a plan to make video analysis with it and for today we wanna test how it will work on GPU load and if it overheats (we need some improvements in heat distribution then). I’m looking for a solution to run some script or benchmark to load GPU as much as possible and see how it works. Problem is that only way to connect to it is SSH connection. I need solution that will work without GUI then. Do you have any idea how to make such a test? We need to load CUDA cores of course.

Thanks for your advice.
Przemysław Chwiała

2

I am not familiar with the Jetson Nano. You might get better answers in the sub-forum dedicated to that platform: https://devtalk.nvidia.com/default/board/371/jetson-nano/

Generally speaking, maximizing the power consumption of a GPU requires a mix of computation and memory activity. What mix is optimal in that regard will depend on the specific GPU used. I sketched a possible solution based on CUDA (a simple program that one can run from the command line) in this thread on Stackoverflow: https://stackoverflow.com/questions/55187907/small-program-that-burdens-the-gpu. It has some parameters, in particular POLY_DEPTH, that allow modifying the balance of computation and memory activity.

3

Thanks, I’ll create topic in sub-forum you linked