Jetson AGX Orin GPU Usage

Hi,

With the source GitHub - rbonghi/jetson_stats: 📊 Simple package for monitoring and control your NVIDIA Jetson [Xavier NX, Nano, AGX Xavier, TX1, TX2] I can see how much cpu and gpu I can use.

Also with source GitHub - theAIGuysCode/yolov4-deepsort: Object tracking implemented with YOLOv4, DeepSort, and TensorFlow. I running the Tracker with YOLOv4-Tiny

FPS Result:

FPS: 8.42
FPS: 7.80
FPS: 8.05
FPS: 8.74
FPS: 8.81
FPS: 7.97
FPS: 9.12
FPS: 8.74
FPS: 8.87
FPS: 8.30

I was expecting higher performance. When I look at the GPU usage I can confirm it is working. But it works at low percentages.

Is there a way to get more out of the GPU? I want to take maximum advantage of Jetson AGX ORIN. I want to do projects such as object detection and object tracking with this device. I’m open to any resources you can suggest.

Best Regards, Hüseyin GÜLEK.

Hi @huseyin.gulek, it is recommended to use TensorRT for inferencing to achieve maximum utilization of GPU, in addition to the Jetson DLA engines (deep learning accelerators). Also for maximum framerate you would want to run the model with INT8 precision, which TensorRT can also do.

Here is kind of a similar project that does YOLO, except it is using TensorRT instead of TensorFlow: https://github.com/RichardoMrMu/yolov5-deepsort-tensorrt

Thank you for feedback,

I will apply it as soon as possible and share the results.