Google's eighth-generation TPUs split training and inference into two specialised chips. Here's how TPU 8t and TPU 8i work, ...