Optimizing speed with Google's Tensor Processing Unit