search

marcoslucianops / DeepStream-Yolo

NVIDIA DeepStream SDK 7.0 / 6.4 / 6.3 / 6.2 / 6.1.1 / 6.1 / 6.0.1 / 6.0 / 5.1 implementation for YOLO models
MIT License
1.4k stars 345 forks source link

YOLO v5 INT8 is not faster than FP16 or FP32 #166

Closed 4399chen closed 1 year ago

4399chen commented 2 years ago

Hello.

I used your project and followed your steps. But INT8 is slower than both FP32 and FP16 in my tests. Test with a video file(640x480x50FPS)

FP32 image

FP16 image

INT8 image

Jetson Xavier NX, MODE 20W 6CORE YOLO v5 6.1 yolov5n.pt

4399chen commented 2 years ago

It is obvious that INT8 is not exerting all the ability of NX. Because when I lower the power of NX, THE FPS of FP16 goes down a lot, but the FPS of INT8 doesn't change much.

FP16(MODE 10W 2CORE) image

INT8(MODE 10W 2CORE) image

marcoslucianops commented 2 years ago

I will try to do some tests this week.

marcoslucianops commented 1 year ago

Please use the updated files, there are a lot of improvements on them.