The 2nd generation of Google's machine learning machine "TPU" appeared, 1 board 180 TFLOPS reached 11.5 PFLOPS at 64 grid

Google will hold a developer conference on May 17, 2017 local time "Google I / O 2017", Specialized machine specialized in machine learning / artificial intelligence (AI) development"TensorFlow Processing Unit (TPU)We released the second generation model of. In the second-generation TPU, we can process not only conventional "inference" but also "training", boasting performance of 11.5 PFLOPS when 64 sets are combined.

Build and train machine learning models on our new Google Cloud TPUs
https://blog.google/topics/google-cloud/google-cloud-offer-tpus-machine-learning/

Google.ai
https://google.ai/

You can check the first generation model of Google's machine learning machine "TPU" in the following article.

It is obvious that Google has made secret the back of a "TPU" dedicated to speeding up machine learning by a few years ago - GIGAZINE


In machine learning, two processes, "inference" and "training" are performed. In the 1st generation TPU, only processing of "inference" was performed, but in the 2nd generation TPU, in addition to inference, processing up to neural network "training" is possible.

This is the 2nd Generation TPU. Four processors are installed in one unit, boasting computing performance of 180 TFLOPS (teraflops) of floating point per unit.


Equipped with a huge heat sink.


It is a 2nd Generation TPU boasting computing performance of 180 TFLOPS with 1 unit, but it is possible to improve the performance by configuring "TPU Pods (TPU Pod)" in grid state by interconnecting with the network.


This is a TPU pod consisting of 64 units TPU that is actually used in Google's data center. The computing performance is wonderful 11.5 PFLOPS (petaflops). Google says that the TPU Pod can apply 11.5 PFLOPS to a single machine learning training task.


In the keynote address of Google I / O 2017, CEO Thunder Pichai said, "When Google began search business, we needed to reconsider the architecture of conventional computing process, built the data center from the place, , As we are required to review the architecture by machine learning and evolution of AI, we are beginning to build the AI ​​First data center again ", with the arrival of the era of" AI First ", AI And TPU as dedicated hardware specialized in machine learning technology. Please note that the 2nd Generation TPU was introduced on Google Compute Engine and is available at the same time as the announcement.


In addition, machine learning techniques including deep learning and the like are bringing about a big breakthrough in various fields such as image analysis for medical use, machine translation technology, automatic game play software and so on, enormous in training of machine learning model In view of the fact that calculation resources are required, in order to accelerate the development speed of researchers,TensorFlow Research Cloud (TFRC)We announced.

Research Blog: Introducing the TensorFlow Research Cloud
https://research.googleblog.com/2017/05/introducing-tensorflow-research-cloud.html

TFRC can access the cloud TPU using second generation TPU that can infer and train processing and will provide the computing performance of 180 TFLOPS and the memory access of ultra wide bandwidth of 64 GB with the same TensorFlow programming interface as before about. In addition, TFRC is not limited to academic research, it is provided to researchers and developers in a wide range of fields with expert knowledge, and researchers / developers can apply for TFRC multiple times with multiple projects OK It is said that.

Based on the idea that TFRC will benefit the entire research community of machine learning, researchers / developers who use TFRC resources can use knowledge gained through machine learning through papers, source code, blogs and open media It is required to publish.

Cloud TPU Alpha program is also available for companies that want to use Cloud TPU for closed research and development, and you can apply from the following page.

Learn more about Cloud TPUs and the TensorFlow Research Cloud - Google
https://services.google.com/fb/forms/tpusignup/