Jetson agx orin executes yolov5s bug

[653576489]

yolov5s fp16 reasoning is slower than fp32

As mentioned in the above link, I found that executing yolov5s half-precision is slower than full-precision on jetson agx orin. The same code is indeed faster than full-precision on jetson xavier nx. It is normal. May I ask? What are the reasons?

[AastaLLL]

Hi,

Which frameworks do you use?
Could you run it with TensorRT?

Thanks.

[653576489]

The code is executed under the pytorch framework. I just tested tensort’s FP32 and FP16 and found that it is normal. fp16 is faster than fp32. What may be the cause of the problem?

[AastaLLL]

Hi,

Have you maximized the device’s performance?

$ sudo nvpmodel -m 0
$ sudo jetson_clocks

More, do you have the ONNX model?
Could you check to get the benchmark data with trtexec?

Thanks.