TPU v3, for example, can train BERT, Google’s game-changing language model, roughly eight times quicker than Nvidia’s older ...