Same Code but get different result on different GPU card

i run a very simple code on Nvidia GPU cards:P4, P100,V100, but get different accuracy seperately: 84%,78% and 77%. After run another time, the different result is still there.
So confused, anybody get similiar issue like this?
below is the code:

sz = 360
bs = 10
path = Path(’./imagenet/’)
ds_tfms = get_transforms(max_zoom=1.02)
data = ImageDataBunch.from_folder(path, test=‘test’, ds_tfms=ds_tfms, size=sz, bs=bs, num_workers=0)

learn = create_cnn(data, models.resnet101, metrics=accuracy)
learn.callbacks = [EarlyStoppingCallback(learn), SaveModelCallback(learn), CSVLogger(learn)]

does this still happen if you pass a seed to from_folder?

i’ll try it later, and try to find any more details