I was developing a CNN with tensorflow in Kaggle with a P100 GPU, it obtained a val accuracy of .87 and a loss of .56. When I downloaded it and tested it on my PC (which does not have a GPU that tensorflow can use) I noticed that Its performance declines, an image that predicted correctly in Kaggle has many errors when predicting on my PC. Why would it be? I think that due to the fact of not predicting images using a GPU, but I would like to know the opinion of someone more experienced

To be more sure what it was, I made a prediction with an image that I took from the training dataset. At the time of training, it obtained an accuracy of .92. The bad thing is that it did not predict it well.