Google’s Tensor Processing Units Change the Rules for Machine Learning

One of the most interesting – and unexpected – announcements Google made at its I/O developers conference last week was that it has designed and implemented its own chips for machine learning. During his keynote, Google CEO Sundar Pichai introduced what he termed Tensor Processing Units (TPUs), saying the company used these in its AlphaGo machines, which defeated Go champion Lee Sedol. “TPUs are an order of magnitude higher performance per watt than commercial FPGAs and GPUs,” Pichai said. While he didn’t give many details, Google distinguished hardware engineer Norm Jouppi explained in a blog post that a TPU is a custom ASIC (application specific integrated circuit). In other words, it’s a chip specifically designed to run machine learning and particularly tailored for TensorFlow, Google’s machine-learning framework.  Image In the…


Link to Full Article: Google’s Tensor Processing Units Change the Rules for Machine Learning