As all of us who are excited about the launch of Oculus and Vive are learning, virtual reality is all about GPUs. While many PCs have enough CPU horsepower and memory to handle a VR workload, very few have GPUs that are up to even the minimum suggested specs for VR playback — let alone development. This nearly insatiable need for GPU horsepower makes VR a natural area of focus for Nvidia, as it showed at this year’s GPU Technology Conference (GTC). With 37 VR-related sessions, dozens of demos, and a good portion of CEO Jen-Hsun Huang’s keynote dedicated to VR, it was, along with deep learning and autonomous vehicles, one of the three biggest themes at the conference.
Much of the excitement around VR is built using amazing demos of virtual worlds. Perhaps fittingly, virtual worlds are actually easier to portray than real worlds in VR, because it is possible to model exactly what a person would see from any point of view, and with any lighting. Creating stereo views is also relatively simple — basic stereo camera support can be added in game engines such as Unity merely by checking a box. Now, though, we’re starting to see full-fidelity experiences based on modeling real-world locations. Huang showcased two of the most elaborate during his keynote: Everest VR and Mars 2030.
Everest VR was put together by Solfar and Rfx using 108 billion pixels of actual photographs, that were turned into 10 million polygons, which in turn drive a VR experience that is essentially photorealistic. Game-like physics are used to generate drifting snow for additional “reality.” What makes immersive experiences like Everest much more powerful than an ordinary 360-degree video is that the user is not limited to a particular location, but can move through the environment.
Similarly, Mars 2030 is largely based on massive numbers of photographs our spacecraft have sent back from there, allowing NASA, with help from Fusion VR, to model eight square kilometers of the planet’s surface. The studio then went to work enhancing the model, including creating 1 million hand-sculpted rocks, and 3D versions of mile-long underground lava tube caves. Hoping for some star appeal, Huang brought Apple co-founder Steve Wozniak up on screen so we could see him be the first ever to experience Mars 2030. Everything went well for the first couple minutes until Woz said he felt dizzy, and needed to stop before he fell out of his chair. That was definitely awkward, and symptomatic of the “queasiness” issue that continues to intermittently trouble VR rollouts.
Nvidia’s iRay is the ray-tracing package of choice for many of the top 3D modeling packages already. This year the company will be extending it with additional capabilities to support the unique requirements of VR. Before VR, producers of computer-generated video content could choose between either time-consuming photorealistic rendering, like that used for feature films, or real-time, plausible rendering needed for interactivity in video games. VR experiences are creating a demand to achieve the best of both — realistic, immersive, experiences that are high-quality, 3D, and allow the user to move around. That means they can’t be entirely pre-rendered. Unfortunately, moving from a 3D model to a photorealistic experience is too processor-intensive to do entirely in realtime. So views of the model need to be rendered, typically using ray tracing to mimic lighting and reflections in a physically accurate way. Traditional applications like movies or print only require high-resolution 2D images to be produced, but immersive VR requires the creation of an interactive experience.
That’s where iRay VR comes in. For those willing to buy, or rent time on, a high-performance computing cluster (like Nvidia’s own DGX-100), iRay VR can generate a navigable model of a scene. The user can move around in the scene, and turn their head, while getting physically-accurate lighting, shadows, and reflections as they move. Nvidia demoed this with an interactive VR model of its planned headquarters that was quite convincing. Unfortunately, even the viewing computer needs to be pretty massive — requiring around a 24GB frame buffer like the one in the 24GB Quadro M6000 to run.
Even with a supercomputer at hand, rendering for VR requires some compromises. Lucasfilm’s Lutz Latta explained that, for example, the Millenium Falcon model used for Star Wars is made up of over 4 million polygons. Perfect for the ultimate in cinematic reality, when it can be rendered one frame at a time, but too complex for a Star Wars VR experience. In addition to simplifying it, the studio has worked on a way to have a unified asset specification, so that models can be built once and then are available for use in a variety of different media like film and VR.
For those of us on slightly more limited budgets, iRay VR Lite will allow you to upload a model to Nvidia’s servers, where they will generate a photorealistic 360-degree stereo view — but one that you can’t walk around. For a full-fidelity experience, even the Lite version will take advantage of 6GB or more of frame buffer, but the experiences will also be viewable on low-end devices like Google’s Cardboard. Nvidia expects to have iRay VR Lite available by mid-year, with iRay VR to follow.
iRay VR is only one part of Nvidia’s VRWorks set of tools for VR development and delivery. Other parts of VRWorks also had updates announced at the show. In particular VRWorks SLI will provide OpenGL across multiple GPUs and there is now a VR support in GameWorks.
Everest and Mars are both very expensive, long-development-cycle, efforts, more or less the equivalent of making a feature film. That limits their creation to large organizations, and their subjects to those that are likely to attract millions. Startup Realities.io has developed a system that allows it to relatively quickly, and inexpensively, create photorealistic environments from “everyday” locations. Using around 350 photographs of a scene, it can use a process called photogrammetry to create an interactive model that the user can walk around.
The level of detail is pretty amazing. You can stoop down and see trash on the floor, or walk over to a wall and see the brush textures in the graffiti. Realities also captures scene lighting, using light probes, so reflections change realistically as you move. If you’re one of the lucky few that have a Vive, you can download Realities for free via Steam.
Even more exciting, Realities founders David Finsterwalder and Daniel Sproll hope to further democratize the process of creating immersive experiences even more by enabling others to go out and capture the images they can then process and produce.
A more common way of creating VR experiences is using 360-degree camera rigs. There are plenty of those on the market, ranging from consumer units with a couple fish-eye lenses to studio-quality rigs like the ones from Samsung and Jaunt VR. However, all of them require a lot of post-processing to accurately stitch the images together. One company, Videostitch, has made a good living providing stitching solutions, but at GTC they announced they’ve gone one step further. They introduced their own turnkey camera + stitcher, the Orah 4i. Available for pre-order for $1,800, it has four high-quality cameras with 90-degree lenses, which feed synchronized data to a small processor box that stitches them and streams a live 360-degree image. In addition to the obvious use for event coverage, the system will also allow excellent real-time previewing for those doing high-end 360-degree production work.
As an aside, the Orah, like most 360-degree camera rigs, does not produce a stereo image. And because it only has one camera facing each direction, you can’t really generate one after the fact. Higher end rigs with more cameras, like Jaunt’s, do allow post-processing to generate depth maps and stereo views.
However, almost all of these units — whether mono or stereo — are bundled together under the umbrella of “VR.” Similarly, many “VR” experiences are static 360-degree photos that don’t allow you to move around the scene (but may or may not be stereo). It’d be great to start to get some standardization on terminology here. In my case, I’ve started to use immersive only for experiences that are both stereo and allow movement within the scene. I think it’d also be helpful if we only used VR to refer to experiences with a sense of depth (e.g. stereo), but the term is too popular as a marketing tool for that to be likely to happen.
VR at the show wasn’t all about entertainment. Two of the demos I experienced were all about commercial applications. ZeroLight was showing off an amazingly realistic Audi in VR — part of a customer-focused sales experience that Audi will start rolling out later this year — complete with Vive headsets in dealer showrooms that allow you to configure and virtually experience your car.
WorldViz has extended standard VR functionality to make it ideal for many industrial and commercial applications. It allows users to see each other and work together on a task in a virtual environment — even if they have different brand headsets — and it supports much larger “room scale” environments. For example, one client created a virtual hospital in a gym, so doctors could test out the work environment before it was built. One of the scenarios I ran through involved working on a helicopter rotor. It really brought home the power of the Vive’s touch controllers. They are a dramatic step forward from trying to use a small remote or gaming controller to manipulate objects in 3D. I hope Oculus gets their version out soon.
VR at GTC was deliberately about nearly every application other than games — after all we just had VR-frenzy at GDC last month — but for most individuals, VR in 2016 will either be about gaming (I’m a racing sim fan, so my favorites that are available so far are Project Cars and Dirt Rally, but there are tons of others) or involve fiddling around with 360 experiences using a Gear VR or Cardboard. For those who do buy a high-end headset for gaming, sure the other experiences are cool, but I don’t see anyone upgrading their computer and plunking down another $800 just to walk around a virtual Mars for a bit. Beyond that, I’m hoping Google announces something amazing under the Android VR banner in May, that can bridge the gap between the current low-end mobile phone offerings and the current crop of gamers-and-hackers-only PC-driven headset offerings.