Google has grand ambitions for artificial intelligence (AI). A large chunk of the web giant’s future business will rely on machine learning to power its range of upcoming services and products like self-driving cars, AI chatbots, and virtual reality devices.
But AI requires powerful computing. While no stranger to building its own hardware, Google has traditionally bought components such chips from already established players like Intel.
This relationship took a twist this week when the company revealed that it has indeed been making its very own long-rumoured chips, designed specifically for machine learning.
Software works best when running on the finest hardware, and that’s why Google said it made the decision to design its own chip.
The entire project, which Google has kept under wraps for a few years, essentially gives the company its own power for machine learning applications. The chip is tailored for TensorFlow, Google’s own open source machine learning software library, and has been up and running in Google data centres since 2015.
“We’ve been running TPUs inside our data centres for more than a year, and have found them to deliver an order of magnitude better-optimised performance per watt for machine learning. This is roughly equivalent to fast-forwarding technology about seven years into the future (three generations of Moore’s Law),” said Google’s Norm Jouppi this week.
“TPU is tailored to machine learning applications, allowing the chip to be more tolerant of reduced computational precision, which means it requires fewer transistors per operation,” he added.
TPUs have been tested already in Google applications such as RankBrain, which is used to improve the relevancy of search results. Street View has also been sucking up power from the TPUs to boost accuracy in Google Maps.
But perhaps the highest profile case of Google’s stealthy use of the chips is in its recent AlphaGo Chinese board game win against Go world champion Lee Sedol.
“Our goal is to lead the industry on machine learning and make that innovation available to our customers,” Jouppi said.
“Building TPUs into our infrastructure stack will allow us to bring the power of Google to developers across software like TensorFlow and Cloud Machine Learning with advanced acceleration capabilities. Machine Learning is transforming how developers build intelligent applications that benefit customers and consumers, and we’re excited to see the possibilities come to life.”
But questions remain over the chip, which was unveiled at Google’s annual I/O conference. It is not yet clear whether the chips will work exclusively with TensorFlow, or will be open to other machine learning libraries. The location of the TPUs’ manufacturing was also not revealed by Google, and seeing as Google doesn’t have its own factory, the job of manufacturing is likely outsourced to a specialist like GlobalFoundries.
Worrying national security concerns for the Britsih government may reportedly scupper Nvidia's controversial acquisition of…