Google in order to further strengthen the machine learning ability, has the secret to develop a dedicated Tensor processor (Tensor Processing Unit, TPU) for several years. Google on the Google I/O of 2016 published for the first time in this chip, and points out that the designed for machine learning design customized chips, far ahead in terms of performance/power consumption than existing solutions for seven years, the equivalent of three cycles of Moore's Law (Moore 's Law).

Google CEO Sundar Pichai said, designed for machine learning application of TPU, in terms of performance/power ratio, than the GPU, the FPGA, and other products on the market at present is obvious higher than, is the company's secret weapon on the development of machine learning. Recently beat human go pro AlphaGo is built, TPU computer operation.

Google's outstanding hardware engineer Norm Jouppi reveal more details information in the official Google blog. For Google, he said, machine learning is the basis of many applications, such as street, the wisdom of the Inbox reply function and voice search. But Google understands that good software on the excellent hardware implementation, results will be more bright, so Google secret in a few years ago began to develop their own machine learning application acceleration, and TPU is the result of the secret plan.

TPU is a customization of application specific integrated circuits (ASIC), designed for machine learning, especially TensorFlow system design. TensorFlow is Google's open source machine learning system. The chip has been used in Google's own data center for more than a year, according to the actual result of the accumulation, according to its performance/power ratio than the average about commercial solution for seven years, leading the equivalent of three cycles of Moore's law.

Because it is designed for machine learning, the chip of computing precision demand for tolerance, less so when performing arithmetic can use transistors resources. Compared with general solution, Google can perform on the self-developed chips more operations, or the use of more complex machine learning model. TPU use customization of the motherboard, through the hard drive slot integrated rack in the center of the data.

It is worth noting that Google to get the first piece of chip test, it took only 22 days to install the chip in psychology in their own information, perform a variety of applications. This shows that Google's chip design team have considerable technical strength.

For machine learning field, Google's goal is to maintain the industry leading position, and take innovation to customers. Through the TPU integrate into their own infrastructure, Google will be for the application of machine learning developers to bring greater acceleration performance.