Two Networks speed different when I apply the two different GPU

vision
1

Hello. Happy New year!!!

I am trying to calculate my suggested network speed and comparison network. Both are semantic segmentation network.
My network name is A.
The other name is B.
The input size of image is same. There is no resize or crop.

From Titan RTX, the network A has 25 FPS. The B has 20 FPS.
In the google colab I get the V100. When I use v100, network A & B has similar speed and sometimes B has better speed. Do you have any idea why it happens?

Thank you.

2

If you are using cudnn, then different kernels might be picked due to the difference in workload (models) and GPU architectures.
You could use torch.backend.cudnn.benchmark = True to let cudnn profile the kernels and pick the fastest one. Note that the first iteration for each new input shape will be slow, since the profiling will be executed.

1 Like
3

Thank you for answering.