“The rise of deep learning is largely due to its reliance on matrix multiplications, which have a favorable ratio of computation (N-cubed operations) to data communication (N-squared bytes).”