Why should data batches in a neural network have an equal size?

I have seen some recent research works on making the batch size dynamic, but still, I can't find an answer to my question.