Google and Movidius Partner to Drive Vision Processing
Everyone’s aware of GPUs, Graphics Processing Units, but most have not heard about (much less begun to use) the next wave in specialized mobile hardware: the Vision Processing Unit (VPU). Movidius, one of Googles Project Tango partners and the company which put VPU System-on-Chip (SoC)s on the map, announced in a post on their web site that they are partnering with Google’s Machine Learning Group.
Google will begin using Movidius’ latest VPU—the MA2450 in its mobile devices and platforms. The MA2450 is the second iteration of the Myriad family of vision processors based on a proprietary software-controlled, multi-core, multi-ported memory subsystem and caches which can be configured to allow a large range of workloads. This flexibility will make it possible to run specialized algorithms for recognizing and tracking highly disparate types of objects at lower computational and power cost.
In a post on TechCrunch about the partnership, Movidius’ CEO Remi El-Ouazzane, is quoted saying that “power consumption savings of the new chip equates to a 10-100x power savings over current models on the market.” In addition, the new chip is 20% the size of equivalent options.
The collaboration isn’t limited to hardware licensing. Google is also expected to implement the Movidius software development environment. As part of the deal, Google’s Machine Learning Group, headed by Blaise Agϋera y Arcas, is expected to collaborate with Movidius’ engineering team on the company’s next generation neural network technology.
Announcements like this serve to remind the market that computer vision is a highly active field of research and development which promises to advance the use of Augmented Reality on wearable devices that must be highly power efficient to recognize and track a vast range of different targets in the real world while serving users for 8 hours or more.