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