Machine learning for many popular Google applications add a lot of charm, currently has more than 100 branches of the Google team in the use of this technology, including street, inbox intelligent response, voice search. During the I / O 2016 conference keynote speech, Google to everybody introduced the recent progress in the fields of artificial intelligent and machine learning, which an interesting detail is the so-called TPU (tensor processing unit) custom chips. In order to help us better understand, Google famous hardware engineer Jouppi Norm specifically in the cloud platform blog published an in-depth introduction to the article TPU.
A TPU board, just can be inserted into the data center rack of a drive bit.
Good software, natural and ultimately excellent hardware blessing. A few years ago, Google has quietly launched a project - to do it yourself, to provide customized acceleration for machine learning applications.
The results of this research, that is, this paper should focus on the introduction of the "tensor processing unit" (Processing Unit Tensor), referred to as TPU. This is a machine learning and customized ASIC, and after the TensorFlow tuning.
TPU has been in the Google Data Center for more than a year, practice shows that it can bring a very good machine learning performance per watt characteristics. TPU is designed for machine learning applications and custom, it is more tolerant, can reduce the accuracy of the calculation (the required transistor operation is also less).
Because of this, second we can squeeze more operation time in the chip, using more complex and powerful machine learning model, faster deployment, users will get more intelligent results more quickly.
TPU has been applied in the "war of the century" with Li Shishi and.
Google is a positive example of TPU from the research to the practical application, the team spent only 22 days time, let it run up in the data center.
TPU has many Google services to provide support, including RankBrain (for improving search results of correlation) and street view (improved precision and quality of navigation map).
In the man-machine chess game with Li Shishi, AlphaGo also has a TPU behind the figure. TPU let AlphaGo be faster thinking, more look far ahead from a high plane before moving on.
Google's goal is to achieve the industry leader in the field of machine learning, while the innovation to provide as many customers as possible.
In the infrastructure of war make TPU, making Google able to developers to bring greater software energy, such as advanced acceleration techniques blessing TensorFlow and cloud machine learning.
Machine learning has changed the way developers build smart apps for consumers, and we're looking forward to it in our daily lives.