Is it expected that the training speed of torch.cuda.amp is better than apex.amp

[kehuanfeng]

The environment is as below,
torch: 1.11.0+cu113
apex: 0.1
cuda 11.3.109
mmdet: 2.24.1
mmcv-full: 1.5.1
model: configs/yolox/yolox_l_8x8_300e_coco.py

I tried to compare the training speed of torch.cuda.amp(autocast + gradscaler) and apex.amp, and found that the native torch.cuda.amp is faster.

# the number indicates the total training time for two epochs,
torch.cuda.amp: 2130 sec
apex.amp: 2309 sec

I'd like to understand whether it is expected and why it happens.

[kehuanfeng]

After breaking down the step time of torch.cuda.amp and apex.amp, it seems like apex requires more data copy.

[ptrblck]

apex.amp is deprecated and you should use the native implementation via torch.cuda.amp as described here.
Closing