Darknet YOLO slowing down when using Jetpack4.4's cuDNN(8.0.0) on Jetson Xavier NX and Jetson Nano

Hi.

In general, FPS increases when using darknet Yolo’s cuDNN(8.0.0). However, in the new Jetpack 4.4, FPS is decreases when using cuDNN.

  1. Jetson Xavier NX
    nvpmodel -m2, jetson_clock
  • Used yolov3-tiny weights
    Not use cuDNN : 45.2 fps
    Use cuDNN : 40.4 fps

  • Used yolov3(320*320) weights
    Not use cuDNN : 8.5 fps
    Use cuDNN : 7.2 fps

  1. Jetson Nano
    nvpmodel -m0, jetson_clock
  • Used yolov3-tiny weights
    Not use cuDNN : 11.3 fps
    Use cuDNN : 6.6 fps

  • Used yolov3(320*320) weights
    Not use cuDNN : 3.1 fps
    Use cuDNN : 2.9 fps

  1. Jetson Nano (Use Jetpack 4.3, cuDNN 7.6.3)
    nvpmodel -m0, jetson_clock
  • Used yolov3-tiny weights
    Not use cuDNN : 11.3 fps
    Use cuDNN : 13.9 fps

I got the same results at https://github.com/AlexeyAB/darknet as well as https://github.com/pjreddie/darknet.

I hope this problem will be solved quickly.
Thanks.

Hi,

We found that darknet run slower with cuDNN 8.0 but the root cause is still under investigation.
https://forums.developer.nvidia.com/t/121579

By the way, our DeepstreamSDK also has YOLO related sample.
It’s recommended to give it a try. Deepstream has optimized for the Jetson device.

/opt/nvidia/deepstream/deepstream-4.0/sources/objectDetector_Yolo/

Thanks.

Hi yacad,

I can only achieve to 7fps when test Jetson Xavier NX with yolov3-tiny weights. I see that you have achieved to 40.4 fps and 45.2 fps. could you give me instruction to achieve that?