I see the readme in this repo does provide the time/epoch for several GPU.
For future reference, here is my result on a Rx 6800, Rocm v5.2.3, Pytorch-Rocm v5.1.1
Time/Epoch:
- Without mixed percision: 2m1s
- With mixed percision: 1m7s
So apparently using mixed percision on Navi 21 does bring up certain improving in training time too.