Here is my code, I just collect images through a USB camera (Intel Realsense D435if) on Jetson Xavier NX.
while True:
# Wait for a new frame
frames = pipeline.wait_for_frames(100000000)
To my knowledge, D435if will finish the ISP process in the camera board, so I think there should not be so many GPU usage during the collection procedure. However, my results shows that the higher FPS I set, the higher average GPU usage it has.
| FPS | GPU Usage |
|---|---|
| 5 | 1.95 |
| 10 | 2.15 |
| 15 | 1.6 |
| 20 | 3.82 |
| 25 | 5.64 |
| 30 | 6.56 |
| 35 | 8.84 |
| 40 | 10.99 |
| 45 | 11.71 |
| 50 | 13.56 |
| 55 | 13.26 |
| 60 | 14.62 |
And I also check the CPU activity by nsys report, as shown in the following screenshot.
Do anyone konws how to NOT use GPU in this process? And why the image collection will use GPU and CUDA? I try to run export CUDA_VISIBLE_DEVICES="", then the collection program will keep blocking.