Unreal Engine and Unity To Get NVIDIA's New VR Rendering Tech (roadtovr.com)

An anonymous reader writes: NVIDIA has announced that Unreal Engine and Unity will see integrations of its new Simultaneous Multi-projection (SMP) rendering tech, which the company says can yield "a 3x VR graphics performance improvement over previous generation GPUs." NVIDIA recently introduced the technology as a unique feature of its latest series of GPUs built on the 'Pascal' architecture. According to the company, Simultaneous Multi-projection allows up to 16 views to be rendered from a single point with just one geometry pass, whereas older cards would need to add an additional pass for each additional view. This is especially beneficial for VR rendering which inherently must render two views for each frame (one for each eye). With Simultaneous Multi-projection built into Unreal Engine and Unity, game creators will have much easier access to its performance benefits. SMP is supported by all of NVIDIA's 10-series GPUs, including the recently announced GTX 1060.

Comments

unreal engine?

[Texmaize]

When I first saw unreal engine, I thought this was going to be a science story about a new reality distortion machine that created a mass hallucination about a certain recent FBI ruling. Sadly, this turned out not to be the case.

Just the latest graphics cards?

[Anonymous Coward]

Is this actually a hardware innovation, or are they just releasing it in the drivers for the latest cards? Hmmm...

Re:Just the latest graphics cards?

[Anonymous Coward]

Yes, this is multiple transforms (up to 16) implemented in hardware at the end of the pipeline for the final view projection(s). This is definitely new hardware.

Re:Just the latest graphics cards?

[Anonymous Coward]

Was this part of the nVidia VR SLI they talked about? Each eye getting a different GPU or something?

Has NVIDIA invented ray tracing?

[Iamthecheese]

Avoiding rendering views that wouldn't show up in the final picture? They call it LMS:

SMP can be used specifically for VR to achieve what Nvidia calls âLens Matched Shadingâ(TM). The goal of LMS is to avoid rendering pixels which end up being discarded in the final view sent to the display in the VR headset after the distortion process.

How is that not ray tracing?

Re:Has NVIDIA invented ray tracing?

[Anonymous Coward]

No, they didn't implement ray tracing in hardware. The LMS technique is described very well here: http://www.roadtovr.com/nvidia-explains-pascal-simultaneous-multi-projection-lens-matched-shading-for-vr/

As you can see, instead of starting with an oversized render followed by distortion, they use the SMP hardware to transform a much closer match to the final render in order to achieve efficiency in the amount of resources required to clock out the final (lens-warped) image to the display.

Re:Has NVIDIA invented ray tracing?

[Guspaz]

How IS it ray tracing? Modern VR requires a lens-correction distortion be performed after rendering so that the image you see through the lens matches what was rendered. Lens matched shading breaks the image up into four quadrants and renders four projected views that are a closer match to where the detail will be in the final result.

Here you can see an example done traditionally:

http://i.imgur.com/FA56wzN.jpg

And here you can see the same scene rendered with lens matched shading:

http://i.imgur.com/CsDouw0.jpg

When a rendered scene normally goes through the distortion shader, most of the pixels around the edges are going to be lost as the distortion is strongest at the edges. This particular technique avoids that by starting out rendering less detail at the edges.

You could also take this to another level and use it for foveated rendering (where you render detail based on where the eye is looking), rather than the current technique of rendering multiple viewports at different resolutions and blending between them. Foveated rendering is a huge win for performance in that it results in a drastic reduction in pixels rendered, but the hardware (very accurate and very low latency eye tracking) and software required to do it isn't quite ready for consumer use.

Re:Has NVIDIA invented ray tracing?

[MobileTatsu-NJG]

Just wanted to thank you for the time you took to write this.

Re:Has NVIDIA invented ray tracing?

[phorm]

"where you render detail based on where the eye is looking"

Just to understand this better, does that mean it has some for of eye/iris tracking and then essentially ups the AA and/or poly count wherever in the scene that is, but renders with less detail around the periphery, similar to how a photo would have a more detail image at the focus point and a more blurry background or surrounding?

I did google it but wanted to be clear from somebody in the know.

Re:Has NVIDIA invented ray tracing?

[Guspaz]

Your eye can only really see detail in a very small area where you are directly looking (in the centre of your vision), but your brain is very good at filling in the blanks and hiding this fact. It drops off extremely rapidly, and for the vast majority of your field of view, you can resolve barely more than basic colour and movement.

The idea behind foveated rendering is, you use eye-tracking to figure out where the user's eye is looking, and then you render a very small full-detail image and place that where they're looking, and then you render a larger lower resolution image and put that behind it, repeating that until you're rendering very few pixels out around the periphery of the vision. Your eye can't tell that the image is getting less detailed (blurrier, really) as it moves away from your centre of vision, because you can't perceive the lack of detail. Obviously I'm simplifying with the progessively-lower-resolution-images description, but that's the gist of it.

When I say the detail drops off really fast from the centre of your vision, I mean it drops off EXTREMELY fast. Check out this graph:

https://en.wikipedia.org/wiki/...

Just 10 degrees off centre and you can only see 20% of the detail. But current rendering systems would still be drawing 100% of the pixels in that area.

To get an idea about how much processing power you can save with foveated rendering in an ideal case, basically consider the area of that graph overall (the whole square) versus just the area under the line. That's a speedup of multiple times.

I'll qualify all this with, I'm no expert, I've just read into it a bit.

Re:Has NVIDIA invented ray tracing?

[phorm]

What would be *really* cool would be to have a 3rd-party display that shows this in effect. If the VR user doesn't notice the difference, but a 3rd-party watching on a screen can see the detail at focal-point and the lack thereof it would be pretty neat to watch.

It would also be good to see how the eyes of different people might work. I'd image the size of the focal point or peripheral detail may be slightly different for various persons.

Re:Has NVIDIA invented ray tracing?

[Guspaz]

There are demos out there you can look at, using modified Rift HMDs. A company called SMI has been working on it. The limitation isn't the understanding of visual acuity, but the overall polish and sophistication of the implementation:

http://www.roadtovr.com/hands-...

Another major issue is the ability to actually derive speed benefits from this approach. If you're implementing it by (as they do in this demo) rendering three different views at different resolutions in different passes, there's a fair bit of overhead involved, and I suspect that they'd also have overlap between the layers where they're rendering more than they need to (can you really tell a GPU to render a donut-shaped view and not spend any time on the pixels in the middle? I don't know, but I'm skeptical)

That, I think, is where nVidia's approach comes into play: by removing the performance penalty of rendering multiple projected views, and using the projection to get the detail (and lack thereof) where you want it to be, basically just a more extreme version of the lens-matched rendering that I linked the screenshots of. Refine that, refine the hardware to the point of being consumer-ready, and you start to see some major benefits.

Advantage Nvidia...?

[Anonymous Coward]

When Nvidia talked on stage about SMP, it sounded more like a proprietary gimmick akin to G-sync (when Freesync is the open implementation that seems to be catching on with more vendors). Now, though, it appears to be useful enough for direct integration into engines themselves. I wonder if this and the GTX 1060 could effectively kill AMDs claim-to-inespensive-VR-fame. It will be interesting to see if AMD can come up with an answer to SMP if it really takes off. Sadly, AMD seems to be lagging this year with just the Polaris 10 and 11.

Its about time

[Anonymous Coward]

So team green finally implemented multi-engine support and are using a foviated rendering scheme with the primary (full resolution) view rendered via the primary engine and peripheral views farmed out to the weaker secondary engines? Congratulations, you've joined the modern console era.

Re:Its about time

[Namarrgon]

Nope, this is used for more efficient VR support in the Unreal & Unity engines. Though the same hardware feature can be used for foveated rendering too.

Isn't this DX10 though?

[Anonymous Coward]

I thought you could do this with Geometry Shaders and RTAs since DX10?

Re:Isn't this DX10 though?

[Khyber]

You could do this with OpenGL 1.0 if you wrote the extensions yourself and had the proper hardware.