“The second generation of Google's TPUs, TPUv2, was designed to handle both inference and training workloads.”