Overview

  • Tensor Processing Unit (TPU)—developed in 2016
  • Since 2012—AI compute power doubling every 3.5 months
    • CPUs and GPUs no longer able to scale
  • Domain specific hardware
    • e.g. Matrix multiplication
  • Faster to train ML models than GPUs
    • 27x faster
    • 38% cheaper
    • e.g. GPU 26hrs vs TPU 7.9mins

Graph View