“In 2012, Google trained a neural network that was 60 times larger than any other at the time, using 16,000 CPU cores.”