Exact same codes, same random seed, huge different outcomes

Hi, I am doing a Resnet50 transfer learning with no pre-trained weights.

For the same codes with the same random state, my friend has both accuracy constantly increasing to 90%, However, my train acc stuck at 40%, valid acc stuck at 33%
We have the exact same codes. The only difference is that he ran the codes using A6000 GPU, mine is the 2060s.

What could be possible reasons for these outcomes?

ps. his num_workers is 2 ,mine is 0 because of windows system, using other number will raise error.

I’d probably check the optimizer and the data-loading setup in more detail.
My usual approach is to have a very detailed log of what is fed, length of the dataloader etc.

Best regards

Thomas