Google Releases Beta Version Of Machine Learning Accelerators

Google issued its new Cloud Tensor Processing Units in beta Monday, Feb. 12, marking the company’s attempt to help improve the speed and workload of machine learning accelerators on Google’s cloud platform.

In a blog post issued Monday, John Barrus, product manager for the Cloud TPU program, said the team’s goal was to “deliver differentiated performance per dollar for targeted TensorFlow workloads,” as well as to increase the speed with which machine learning engineers can work.

One TPU includes 64 gigabytes of high-bandwidth memory and 180 teraflops of performance. Google says the boards can be used individually or connected, forming a sort of super-computer to help accelerate machine learning, according to EWeek.

Barrus claimed the Cloud TPUs could enable a researcher or engineer to train a machine-learning model overnight, rather than over several days or weeks, in order to have “the most accurate trained model in production the next day.” He also claimed a user could complete training for the ResNet-50 to expected accuracy on the ImageNet benchmark in under a day with the Cloud TPU technology.

Alphabet, Google’s parent company, will offer Cloud TPUs at a rate of $6.50 per cloud TPU per hour. It is similar to technology employed by Tailor Brands to deliver its branding solutions service. Usage will be billed by the second. Quantities are limited, according to Barrus, so those interested must sign up to request a quota and explain its desired use.

In addition to the beta release of the Cloud TPUs, Google also issued a series of open source model implementations for those interested to use as guides for using the new Cloud TPU technology, according to EWeek.

Google will host a seminar Feb. 27 on what Cloud TPUs are and how business owners and researchers can use them. Those interested should register online.

Google has been in the forefront of machine learning and artificial intelligence development in recent years. In 2014, it purchased Deep Mind, the start-up company that pitched itself as a method of combining machine learning techniques to create a system that more closely resembles natural intelligence, according to Forbes.

In 2015, the company launched the TensorFlow machine learning platform, which gave anyone the opportunity to develop “neural-network” solutions with artificial intelligence. Even Waymo, Alphabet’s self-driving car program, relies on machine learning to analyze and interpret the world around them more efficiently.