GPU usage discussion

[Chotipon]

I'm working with yolov8 for detection task, from about 10 cameras source were input continuously. And the question is "Is it possible to reduce gpu usage to be about 10% ? " and "How to, if it possible" , i have tested on my 3060Ti which utilization is peak at 90-100% on yolov8n.pt model.

[glenn-jocher]

Reducing GPU usage to about 10% while handling inputs from 10 camera sources can be challenging due to the computational demands of detection tasks, especially with a powerful model like yolov8n.pt. Here are some suggestions you might consider:

1. Reduce Model Size: If you're not already, use a smaller model variant like YOLOv8n-lite which could reduce GPU load.
2. Lower the Frame Rate: Process fewer frames per second from each camera if real-time processing isn’t essential.
3. Decrease Image Resolution: Lower the imgsz during detection which reduces the amount of computation required.
4. Batch Processing: Combine frames from several cameras into batches to optimize GPU usage during inference.

Here’s a quick example of using a smaller model and reducing image size, which could help lower GPU utilization:

yolo detect predict model=yolov8n-lite.pt source='camera_input_path' imgsz=320

You might need to experiment with these adjustments to find a balance between performance and resource usage that suits your needs! 🚀