Deep learning with CPU much lower accuracy than GPU?

Hi everyone,

I've been wracking my brain for the past several days trying to understand why training my CNN on a CPU yields much lower accuracy than on a GPU. Then, I realized that implementing the MATLAB example (https://www.mathworks.com/help/vision/ug/semantic-segmentation-using-dilated-convolutions.html) with the MATLAB command also performs much worse when trained on a CPU vs GPU. What's going on here?

Here are some screenshots of the results on my system:

Thanks!

[Adam]

You haven't given many details. Are you training for the same number of epochs/iterations in both cases?

[Sabina Stefan]

Yes! I ran the example exactly as it is without changing any parameters (same for my own data).

[Adam]

I just ran both on that example and got slightly higher accuracy with GPU than CPU, but given randomness not enough to conclude it is generally better, but certainly enough to conclude that CPU wasn't strictly better than GPU. You on;y posted one set of results above. which were they and what did the other look like?