Jeremy clarifies it more in lecture 3:
‘precompute = True’ caches some of the intermediate steps which we do not need to recalculate every time. It uses cached non-augmented activations. That’s why data augmentation doesn’t work with precompute. Having precompute speeds up our work.