Every year, one of the important event of Google I/O developer conference in 2016 in the United States time Monday morning, Google set out a lot of new things, this time in detail to see parent reports of science and technology, there is a notable thing, Google revealed at the conference company is building a machine learning dedicated chip, TPU (Tensor Processing Unit).

Google ai Alpha Go in has won the south Korean Go nine masters 乭, lee AlphaGO the key to winning, lies in the combination of Machine Learning, Machine Learning) and deep neural network (deep neural networks), the simulation game master how to analyze disk, find out the best moves, step-by-step. Google CEO sander, Ms. Pichardo (Sundar Pichai) on the 18th, Google I/O of further revealed the secret.

Pichardo, points out that Google's specially designed for deep neural network for special applications of IC (application - specific integrated circuit, ASIC), based on ASIC hardware or software, will be able to through the analysis of a large amount of data to study the special mission, Google by neural network to identify objects, the face in the photo, know, speak to Android phones instruction, and the translation techniques, even change the Google search engine. Is also the technology improve the computing speed of the Alpha GO, thought and make it look more profound.

Google will use this technology to build a machine learning dedicated chip called TPU (Tensor Processing Unit), Google said in its own blog articles, because TPU run solely for machine learning, to more traditional CPU, GPU, lower precision, the number of transistors in the calculation, can reduce the nature, and therefore, can be squeezed more efficiency, from the transistors per second to perform more complex and powerful machine learning module, and accelerate the use of modules, makes users get the answer faster, Google points out that the team has been in the data center perform TPU for more than a year, and found that the TPU can let the machine learning per watt improved one order of magnitude, roughly speaking, the equivalent of Moore's law microtek piece performance forward for seven years or three generations. Carrying TPU machine board as long as installed in a Google data center on the hard drive slot of can.

TensorFlow Google published last year in machine learning system, and emphasize its open source, means that anyone can use or modify the software engine, but does not mean that the design of the external users can share the Google TPU, but can through Google cloud service using Google all kinds of machine learning software and hardware service.

And Google revealed the project also alerts for semiconductor giants, Google did not use, or even less use chip with IC companies, but its own research and development, Microsoft is now using FPGA to accelerate the artificial intelligence operations.