In-Datacenter Performance Analysis of a Tensor Processing Unit