A domain-specific supercomputer for training deep neural networks

Google's TPU supercomputers train deep neural networks 50x faster than general-purpose supercomputers running a high-performance computing benchmark.