TPU v8

Multi-GPU Kernels, Intelligence per Watt, Heterogeneous Inference, and More | YC Paper Club

Y Combinator

Mentioned as an example of early chip specialization in training data centers.