Google announced on Wednesday, May 18, at the Google I/O conference event, that it designed its own computer chip for deep neural networks.
According to Wired, CEO Sundar Pichai said that Google has designed an application-specific integrated circuit (ASIC) aimed to drive deep neural nets, an artificial intelligence (AI) technology that is reshaping the Internet. These are networks of software and hardware that analyze vast amount of data in order to learn specific tasks.
Google uses neural nets to recognize voice commands in Android phones, identify faces and objects in photos, or translate text from one language to another. Even the Google search engine is transformed by the applications of this new technology.
Because it underpins TensorFlow, Google's software engine that drives its deep learning services, Google calls its chip the Tensor Processing Unit (TPU). TensorFlow was released by Google this past fall, under an open-source license. Any developer outside the company can use and modify the software program.
Google said in a blog post that TPU is tailored to machine learning applications, requiring fewer transistors per operation and being more tolerant of reduced computational precision. Google has not shared the designs for the TPU, but outsiders can use the company's machine learning software and hardware via various cloud services.
According to Computerworld, the new TPU built by Google is a major leap forward for intelligent applications. By creating its own custom chip, the company has taken a big leap forward with the speed of its machine learning systems.
Google has been testing the TPU for over 1 year. Based on job ads posted in recent years, Google was rumored to have been designing its own chip, but until today it did not make its project public.
Other companies have also incorporated deep learning into a wide range of Internet services, including Twitter, Microsoft and Facebook. The neural nets are typically driven with graphics processing units (GPUs) made by companies like Nvidia. But some companies are also exploring the use of field programmable gate arrays (FPGAs) that can be programmed for specific tasks.
