“Google's TPUv1 was 15 to 30 times faster and 30 to 80 times more energy-efficient than contemporary CPUs and GPUs for neural network inference.”