SoylentNews

takyon writes:

Google will include an additional processing chip in its future mobile devices (such as the Nexus line of smartphones and tablets) to enable "deep learning" applications without (necessarily) communicating to a central server:

Google has signed a deal with Movidius to include its Myriad 2 MA2450 processor in future devices. The search giant first worked with Movidius back in 2014 for its Project Tango devices, and it's now licensing the company's latest tech to "accelerate the adoption of deep learning within mobile devices."

[...] More recently, Google managed to cram a [neural] network into its Translate app, allowing users to convert the text in images on the fly. And SwiftKey also runs a small-scale network for word predictions in its SwiftKey Neural application. But all these applications require a large amount of processing power for what are relatively inane tasks. That's where Movidius' chip comes in.

The Myriad 2 MA2450 is referred to as a "vision processing unit." It's really got a single purpose: image recognition. The architecture has very little in common with a traditional CPU, and it's designed specifically to handle the myriad (get it) simultaneous processes involved in neural networks. As such, its power draw when, for example, recognising a face or an image, is much, much lower than doing the same task with a Snapdragon processor. As for how exactly will Google utilize the chips, that's something we're unlikely to know until it's ready to announce devices.

TechCrunch has some additional details about Movidius MA2150 and MA2450 chips. Or look at this product brief (PDF).

Original Submission