HoloLens V2 Will Have A Custom-Made AI Coprocessor
The AI invasion isn't coming. It's already hear. Forget Skynet, we already have Siri, Cortana, Google Assistant, and, now, Bixby. But AI is found in more than those. It's also present in Google Search, Office Lens, email, photos, and, soon, your head. No, we're not talking about plugging your head to a computer, but something close. Microsoft researchers have just revealed how the next version of the HoloLens will include a custom-made, special-purpose chip whose processing power will be dedicated solely to, what else, AI.
What need would an augmented-reality heads up device (HUD) like HoloLens have for AI and machine learning anyway? Unlike virtual reality, AR has to take into account real-world objects in addition to virtual ones. As such, it needs to be able to process and identify such objects, which is one of the primary goals of computer vision. And computer vision, in turn, has recently been applying deep neural networks (DNNs) and deep learning to make huge strides forward.
The problem for both Microsoft's Research team as well as its Holographic team isn't the theory. It's the hardware. Chip makers like Intel, AMD, NVIDIA, and Qualcomm have all boasted of new processors catering to AI and machine learning. Microsoft, however, thinks those aren't enough. These processors are all general purpose ones that, while powerful, aren't exactly designed specifically to address the massive parallel computing required for deep neural networks. So Microsoft's solution is to make its own.
So for version 2 of the HoloLens, Microsoft will put an AI coprocessor inside, together with the already existing HPU. THe Holographic Processing Unit is already responsible for processing masses of data coming from various sensors. The still unnamed AI coprocessor will only concern itself with processing data for deep neural networks and nothing else.
Having a separate and custom-made coprocessor has two advantages. One is that it allows Microsoft to build silicon specifically suited for DNNs without having to worry about general purpose computing uses. For another, it also means that the HoloLens won't have to offload such processing to external computers or the cloud, keeping it a self-contained HUD. What it eventually means for battery life and heat, however, is something Microsoft has yet to iron out. It doesn't seem to be in a rush, however, since HoloLens V2 is still far away.
SOURCE: Microsoft