Suddenly, Intel has been quite busy on the VR/AR/MR front. At IDF 2016, the company announced and outlined multiple facets of its VR plans, including its Project Alloy VR HMD and new RealSense 400 Series camera, as well as plans to develop specifications for mainstream HMDs--all of which involves partnering with Microsoft. Intel quietly demoed a couple of intriguing proofs of concept, and now it's made a couple of key acquisitions to bolster its deep learning IP, which actually pertains directly to its VR/AR/MR vision.
But those are the ingredients, and there’s a larger soup that Intel has cooking. We’ve dug deep to suss out what we do know, what we think we know, and what we suspect. Two of the most important components here are the two types of HMDs that Intel is developing, and Microsoft’s role in both.
Project Alloy VR HMD
To be clear, Intel did not demo its Project Alloy VR HMD at IDF. Yes, there were onstage demos at the keynote, but the HMD wasn’t even using the newer version of the RealSense camera that it will eventually employ, and there were no further opportunities to see Alloy in action.
Project Alloy, in whatever stage of completion it’s in, is different from Rift/Vive and HoloLens. Although I chafe at this seemingly irresistible need companies seem to have to invent new terminology for their technology, I see what Intel is going for when it called Project Alloy “Merged Reality.” Although the HMD is fully occluded like the Rift and Vive, there’s a RealSense camera mounted on the front that scans the real world in (essentially) real time and reproduces that imagery inside the HMD. In that sense, you can “see” the real world. It’s effectively a passthrough system, but whereas the HoloLens has clear lenses that let you see the real world, Alloy reproduces a facsimile of it for you.
Further, Project Alloy is a completely self-contained system. That is, there is a PC crammed inside the HMD, and there are no external wires. We don’t have full specifications--or really any specifications--but we know that it will run on an Intel Core i5 or i7 Skylake CPU and that there will not be a discrete GPU. It also has what a presentation slide referred to as a “sensor coprocessor,” which we believe is just another term for a vision processor (VPU). We can’t be certain at this point, but we believe that, in Alloy, this chip is (or will be) made by Movidius.
It’s highly unlikely that Alloy would employ Microsoft’s HPU (a proprietary term for that specific vision processor), because that chip was custom-made for the HoloLens, and further, an Intel-made device would likely sport a VPU from its new acquisition, Movidius. Whether Alloy might have a Movidius Myriad 2 VPU or a new, unannounced chip is anyone’s guess at this point.
It will rely on Windows 10 for its OS, so it’s an actual PC, and what’s more, it will make use of the Windows Holographic shell. Clearly, then, Windows and Intel are working closely together on Project Alloy.
On this point Intel is being a little cagey, but it did announce that Windows Holographic will run on “mainstream” PCs when it’s available in 2017. (As a teaser, Intel demoed a merged reality experience at 90 fps on a NUC.) But, in any case, the PC that comprises the business end of Project Alloy is decidedly “mainstream,” and doing the arithmetic, you can imagine that Project Alloy will be able to run any of the experiences that a HoloLens can. (Remember, the HoloLens is also a self-contained system, but it has a less powerful Cherry Trail SoC.)
To understand further how Project Alloy compares to other HMDs out there, here’s a simple but handy chart:
|Rift/Vive/OSVR||Sony PSVR||Microsoft HoloLens||Intel Project Alloy||Mobile VR (Gear VR, Daydream)|
|Type||VR||VR||AR/”Mixed Reality”||VR/AR/”Merged Reality”||VR|
|System||Separate PC||PlayStation 4||(Self-contained)||(Self-contained)||Separate smartphone|
This is where the first big head-scratcher lands: If Project Alloy is a fully occluded headset, how and why is it using Holographic technology? The whole idea behind holograms is that they’re placed within the real world (i.e., augmented reality). Further, the display technology HoloLens uses has to do more with projection than anything else, and that would seem to be completely at odds with a fully occluded HMD like Project Alloy.
On the other hand, putting low-GPU-intensive holograms into an HMD’s display should be pretty easy, provided the HMD has a virtual space in which to put them. On top of that, Intel and Microsoft have been clear that their collaborations will allow both 2D and 3D applications within a given XR environment, including those that already exist within Windows. (More on that later in this article.)
Further, as I mentioned, this is not merely an occluded HMD. It gets effective passthrough to the real world by way of the RealSense camera. And that’s where Project Alloy gets really (really, really) interesting.
RealSense 400 Series And Deep Learning
Just as we know few details on the internals of Project Alloy, we know little about Intel’s latest RealSense camera, the 400 series. I speculate that Intel hasn’t finalized it yet; the onstage demos at IDF used an existing RealSense camera, and even during the RealSense session later in the week, we never got any specifications on the new kit.
In any case, this new RealSense camera promises extraordinary capabilities. In an IDF presentation, Dr. Achin Bhowmik, VP and GM of Intel’s Visual Perceptual Computing Group, laid it out as simply as possible: The goal of RealSense technology is to mimic human perceptual systems--both human vision and the human vestibular system.
That is an incredibly lofty goal, but the Group is working towards it, and latest iteration, the RealSense 400 Series, is a step closer. According to Dr. Bhowmik, the new camera will have increased accuracy over its predecessors, will work both indoors and outdoors, offer a smaller and lighter package, and will have a flexible design so that it can be deployed in a number of different types of applications.
Paired with a device like Alloy, the RealSense 400 Series camera and its VPU will offer 6DoF hand and body virtualization/tracking and 3D scanning of objects and spaces, such that you can interact with them and avoid bumping into them (unless you want to). Unlike the HoloLens, which has to scan and “learn” a room before it can place objects within it, RealSense promises to perform multi-room scale movement and tracking in near real-time. It should also enable multiple people to enter the same virtual space, and let individuals interact with holograms.
Intel has been banging away on these capabilities for a while. A year ago at IDF, we saw a (pretty rough) demo of a Leap Motion camera strapped to an OSVR HMD. Intel has moved on from both, replacing the Leap Motion camera with its new RealSense camera, and creating its own HMD instead of using OSVR. Earlier this year at Mobile World Congress, we saw Intel demo a Project Tango smartphone that could perform hand and object virtualization. (Not “tracking,” but “virtualization.”) They had a mocked-up HMD for it, too, and this demo was far more impressive.
It’s all culminated with the announcements of Project Alloy and the RealSense 400 camera. However, powerful though the RealSense camera may be, the most important aspect to it is the presence of deep learning.
A full discussion of what deep learning is and how it works is far beyond the scope of this article, so for more in-depth reading on the subject, check out this article on the rise of client-side deep learning (which is germane to the RealSense 400 series camera and Project Alloy).
Deep learning is a component of artificial intelligence (AI), an area into which Intel is aggressively pushing. But how that pertains specifically to Project Alloy and the RealSense 400 series camera is more simple. In a (very small) nutshell, there are two components to deep learning: “training” and “inference.”
The training portion requires extraordinary resources. This is Big Data, supercomputing country. However, the inference piece can be done with extremely low compute power. Movidius, for example, has its Fathom neural compute stick, which we affectionately refer to as “deep learning on a stick.”
A device like the Fathom (or a chip that fits into such a tiny form factor) can be loaded with a neural net, which may take up just dozens of megabytes of space. The neural net has been trained so that when the device “sees” objects, people, animals and spaces, it can interpret what they are--not just that something is a dog, but that it’s a Golden Retriever, for example.
What’s truly extraordinary about devices equipped with neural nets is that they don’t need any cloud connection to work. They’ve already been trained. However, the neural nets are also learning as they “see” the world, and they can send that data back to to the cloud to contribute to better training. Then, armed with more data, the training part gets smarter, and you can download a smarter neural net to your device, and so goes the virtuous cycle.
With Movidius in its stable, Intel has both parts covered--at IDF, the company announced that it acquired Nervana Systems, a deep learning/AI company.
On an HMD like Project Alloy, where a neural net could be paired with the RealSense 400 Series camera, the potentialities are mind-boggling.
An Unnamed Mainstream XR HMD (Or HMDs)
Lost in the bluster of the Project Alloy excitement is the fact that Alloy is not the only HMD project Intel is working on. Intel representatives would not discuss many particulars, but based on multiple conversations we had with them, we’re able to take some clues and piece them together.
Fundamentally, Intel is looking to address more than one HMD market. Project Alloy is a self-contained system with passthrough capabilities afforded by real-time mapping and tracking. The other unnamed pursuit here is a mainstream-level, fully occluded HMD designed to work with PCs and Windows--although language from Microsoft as well as breadcrumbs inadvertently dropped by Intel reps indicate that an augmented reality device like the HoloLens could be in the offing, too (I find it striking, incidentally, that Intel hasn’t even named this second Thing yet. There’s no project codename. It’s just an idea, a strategy, a notion.) And in this unnamed effort, Intel and Microsoft are working closely together.
This other headset idea will ostensibly fit the mold of Rift or Vive more than Project Alloy does, insofar as it will be an occluded VR HMD designed as a PC peripheral, but it will be a decidedly more mainstream device. Intel’s sees a huge hole in the XR market--that vast unsettled territory between mobile devices like Gear VR and high-end HMDs like Rift and Vive.
Obviously, questions immediately arise as to what kind of quality one could expect from such a device, but Kim Pallister, Director of Intel’s Virtual Reality Center of Excellence, said, “You need more affordable solutions, and that doesn’t mean dumbing down the specs, because then you get a bad experience. You have to re-engineer how the pipeline works to a degree.”
That’s easier said than done, but Intel is looking to solve the problem by leveraging technology it knows best: processors.
Although we don’t have many specifics yet, we do know that Intel is working on solving problems in XR by offloading certain tasks to a CPU. Intel engineers have been working on a version of this with some VR game devs on existing PCs, to alleviate some of the GPU’s overwhelming VR responsibilities. It appears that from this experimentation, Intel is drawing up ways for HMDs themselves to offer similar assistance.
“In addition to having some sensors on the HMD and doing some of the sensor fusion-type work, there’s a theory we have that we’re going to go do a proof of concept and try and prove out that there may be advantages to doing things like barrel distortion, and time warp, and chromatic error correction on the HMD itself,” Pallister told us. “Because you can basically have those close to the sensors and be very responsive in terms of how late you latch the sensor data, and how quickly you update the timewarp, and you can do things like decouple that from the resolution and framerate of the application,” he added.
He further noted that some of the other advantages to this decoupling include rendering the back buffer and upscaling. “You can maybe update the HMD very quickly and time warp but have the simulation running at a different framerate,” said Pallister. He qualified Intel’s results thus far as “pretty good” and believes that the company can continue to extract more VR quality from lower-power hardware.
One of the obstacles in Intel’s way, though, is the need for Microsoft’s help.
Intel’s Daydream With Microsoft
“This collaboration with Microsoft has two components to it,” said Pallister. “One of which is working with [Microsoft] to get to a set of specifications for PCs and for head-mounted displays that will match up and provide mainstream solutions--we use that word very deliberately--mainstream solutions for this Windows Holographic Windows Shell, and a set of APIs that will enable VR for a volume market on PCs.”
In other words, Intel is not looking to make a single, mainstream-level VR HMD. It’s trying to jumpstart a whole new ecosystem, and it’s working with Microsoft to get the software and OS pieces to jibe with the hardware. Then, Intel will provide hardware to other companies, who will make end user products.
This is, you will notice, exactly what the chipmaker has done for years in the CPU market. Just as you can’t buy an Intel PC, you won’t be able to buy an Intel VR HMD. So why is Intel building HMDs?
Pallister was quick to deflate any notion that Intel is getting into the headset business. “We built prototypes, proofs of concept and things like that, but that doesn’t mean we’re necessarily getting into that market,” he said. “As things stand right now, we’re far more interested in enabling the industry...than trying to get into a vertical business.” He noted that this prototyping is designed both for Intel’s engineers to understand what the system-level requirements will be for their Core processors and graphics and so on, and also to provide sufficient knowledge to its industry partners.
He added, “Part of why we’re doing this collaboration with Microsoft on both of these classes of devices [meaning Alloy and mainstream VR HMDs] is that VR is a very high performance, latency-sensitive usage model, so it’s not enough to just say, 'Hey we’ve done this layer, and you figure out the rest.' You kind of have to provide a really tight solution and then work from there.”
An apt comparison here would be what Google is attempting to do with its Daydream VR platform. Google is not interested in making HMDs. Daydream is a platform that Google is providing that OEMs will ostensibly use to build their own VR devices.
Pallister agreed that the above is a more or less sound way of thinking about Intel’s efforts in XR, then clarified with a light chuckle, “Substitute the PC ecosystem for the phone hardware ecosystem, and substitute Windows and Windows Holographic for Android and Daydream.”
In both cases, the goal is the same: to democratize, or even commoditize, XR in such a way that it’s a high-quality user experience that is also affordable and easy to acquire.
But that’s just the first couple of steps--making the stuff work. As we’ve seen time and again in the tech industry in recent years, content, or the lack thereof, is of utmost importance. To that end, though, Pallister seems wholly unconcerned, because assuming the Wintel bit works like it’s supposed to, the content already exists. The answer for the unnamed mainstream mystery HMD is the same as the one for Alloy. “Similar to what [Microsoft] did with HoloLens, they’re giving you a way to bring your existing Win32 apps and content into a VR space and interact with them in some ways. They’re also bringing the set of content they’ve done for Windows Holographic, and for the HoloLens, into the occluded headset Windows Holographic shell as well, so that’s another pool of content that they have to draw from.”
Although the everything-you-already-have-on-Windows is a compelling potentiality, it’s one we’ve heard before, and it hasn’t worked out especially well, at least not yet. (Windows Mobile/Continuum, we’re looking at you.)
The first version of the Intel/Microsoft-specification for these HMDs is expected in December.
So What’s The Grand Plan?
Taking all of the above into consideration, what does this XR utopia actually look like? Despite everything we know, Intel’s XR strategy frankly remains hazy.
Intel is not going to build and sell Alloy HMDs. Intel is building an Alloy platform that other companies can use as a basis upon which to develop shipping products.
Intel is not going to build and sell mainstream VR HMDs that are effectively lower-end versions of Rift and Vive. Intel is going to develop technology and possibly create prototypes to help develop IP with Microsoft that other companies can use as a basis to develop shipping products.
In both cases, Intel and Microsoft are an item, as it were.
Perhaps a helpful way to understand all of the above is to contrast Intel’s approach with that of a company like Oculus. It is true that Oculus has huge VR plans, but they center squarely on its Rift HMD. It starts with a single product; continues with a platform, a catalog of titles, and additions like the Touch controllers; and ends with--well, no one knows what it ends with yet. In any case, Oculus’ strategy of starting with a Thing and growing from there is what you could call an inside-out strategy.
Intel is taking on VR with an outside-in approach. Instead of focusing on one product, it’s building up a massive portfolio of IP--including RealSense cameras and deep learning--that could be leveraged for numerous aspects of XR.
This is exactly what Intel has done in the PC market: Be everywhere. Be in everything.