Figure 1: during the 2016 Google developer Conference I / O, CEO Sandal
In July 26th, according to foreign media, Google was no longer satisfied with the development of AI chips for home data centers, which is now being designed to integrate AI chips into products produced by other companies.
2 years after the release of Tensor Processing Unit (TPU), Google launched Edge TPU on Wednesday local time in the United States, which will enable sensors and other devices to process data faster.
This chip can be used in a variety of scenes, but one of the original uses is in the industrial manufacturing field: consumer electronics manufacturer LG is testing the chip in a system that detects the manufacturing defects of the glass for the screen.
In 2017, Google said its AI chip is becoming more strategic. In the field of AI, researchers are using large amounts of data to train models so that machines can predict when new data arrive.
The original version of TPU can only make these predictions, and the second version (released in 2017) can be used to train the model, which enables it to compete with the Nvidia graphics card. The third generation of TPU was released earlier this year.
Now we have the Edge TPU, which is a microchip specially designed to deal with the AI prediction part, which is less computable than the training model. Edge TPU can run its own computation without having to connect to a number of powerful computers, so applications can work faster and more reliably. They can work together with standard chips or microcontrollers in sensor or gateway devices to process AI.
Injong Rhee, former chief technology officer of Samsung, said Google did not let Edge TPU compete with traditional chips, which would be very beneficial for all silicon chip vendors and equipment makers to be.Edge TPU
Google is not the only cloud computing service provider that is interested in the so-called Internet of things. The core of the Internet of things is to manage and handle data from many small embedded devices. Earlier this year, Microsoft announced the design of its Internet of things chip. Google's new chip will run a model based on the simplified version of TensorFlow AI software, which was released by the open source license in 2015.
The CNS team that LG is responsible for helping internal and other companies handle IT services has been testing Edge TPU and plans to start using them to check the equipment on the internal production line.
At present, in the process of producing glass for display panels, the detection device can process more than 200 glass images per second. Hyun Shingyoon, chief technology officer of CNS team of LG, said that any problems need to be checked manually, and the accuracy of the existing system is about 50%. And the accuracy of Google AI can reach 99.9%.
Hyun Shingyoon also said:
Google has built toolkits, including Edge TPU, NXP chips and Wi-Fi connections, for trial by developers. The company is working with Arm, Harting, Hitachi Hitachi (Hitachi Vantara), Nexcom, NOKIA (Nokia) and NXP manufacturers.
Injong Rhee did not disclose whether Google plans to build a more powerful Edge TPU for training models. (compiling / golden deer)