Over the last couple of years, the idea that the most efficient and high performance way to accelerate deep learning training and inference is with a custom ASIC—something designed to fit the specific ...