Wonder if VR could enable GPUs with less RAM. If you take the GTX 980TI, lets say they get 100 good dies per wafer and the 28nm wafer is 5k$ so one chip is 50$ )excluding testing and packaging and all that) while at the same time maybe they pay 60$ for 6GB of GDDR5 (disclaimer - this estimate might be way off). The board and all that is on it, the cooling add a bit more but the bulk of the costs are the GPU and the RAM ,with the RAM being a huge chunk nowadays. Next year VR vs 4k should require less RAM since VR is lower res and higher FPS, unless some of these VR specific features are memory heavy. Any thoughts on the viability of going maybe more on die cache and less RAM for VR?
Seems feasible. I think we are in a VRAM spoke right now with the new consoles and 4k become more reasonable, but as you say VR focussed on higher shader performance and frame rate, not massive textures and pixel count.
If anything, faster memory like HBM will be needed, not increased capacity.
I thought the same thing. Seems the idea though is that the pixels on the edges are squished, hence after the distortion, the edges of the screen are actually higher resolution than the centre. With this technique, the resolution at the edges of the screen could be slightly lowered, meaning after the distortion is applied which squeezes together the edges, the end result should be an optically consistent resolution across the entire screen. This is all in theory. It honestly seems more effort than it's worth.
Pretty sure it's good practice to render at a slightly higher res than native to keep the center "blown up" area looking nice post warp. This leads to even more waste on the outer areas. Still, it's probably not worth the effort.
Exactly: they're not looking at making the edges blurry, they're removing information which couldn't be seen there anyway due to the lens warping.
"This is all in theory. It honestly seems more effort than it's worth." In many cases people would go to far larger efforts for even a 10% improvement in shader performance. Unless a developer doesn't care about performance and optimization at all.
In a VR headset quite a lot of the screen is not actually visible. For reference the automap, health, shields, and ammo displays in the borderlands games are outside your vision no matter where you point your eyes. Everything is being super-sampled for not just the distortion correction but also since aliasing is very obvious when a screen is magnified by the lenses. The lenses have a sweet spot of clarity and the image also blurs the further you get from it. A FOV of 120 must be set in games to display correctly as well. So the demands of consumer VR with the release of the Vive and Rift is a 2160x1200 supersampled for distortion correction with a 120 FOV, stereo rendered, and a framerate of 90 fps solid to match their 90 hz display so the head tracking doesn't stutter and make you sick. Any performance saving we can take by not rendering high resolutions in the out of sweet spot zones and not rendering at all in the areas invisible to us is much needed in VR. This type of rendering will make higher FOV headsets in the future worthwhile from a performance stand point. I will illustrate the current position of performance without this on the Oculus DK2 which is currently available. The DK@ is only 1920 x 1080 and needs to maintain 75 fps for 75 hz. With a GTX 970 I have to turn all the settings in Bioshock Infinite to medium/low for Geometry 3d in the DK2. Not typically a demanding game by any means. We have all grown accustom to a certain level of graphical fidelity. Having to drop your settings so drastically for VR is sure to disappoint and this Multi-Res Shading is going to mitigate that somewhat.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
9 Comments
Back to Article
jjj - Sunday, May 31, 2015 - link
Wonder if VR could enable GPUs with less RAM.If you take the GTX 980TI, lets say they get 100 good dies per wafer and the 28nm wafer is 5k$ so one chip is 50$ )excluding testing and packaging and all that) while at the same time maybe they pay 60$ for 6GB of GDDR5 (disclaimer - this estimate might be way off). The board and all that is on it, the cooling add a bit more but the bulk of the costs are the GPU and the RAM ,with the RAM being a huge chunk nowadays.
Next year VR vs 4k should require less RAM since VR is lower res and higher FPS, unless some of these VR specific features are memory heavy.
Any thoughts on the viability of going maybe more on die cache and less RAM for VR?
TallestJon96 - Monday, June 1, 2015 - link
Seems feasible. I think we are in a VRAM spoke right now with the new consoles and 4k become more reasonable, but as you say VR focussed on higher shader performance and frame rate, not massive textures and pixel count.If anything, faster memory like HBM will be needed, not increased capacity.
invinciblegod - Sunday, May 31, 2015 - link
This assumes that you are always looking straight ahead right? Does that mean you can turn your eyes or else blurry image?phantomferrari - Sunday, May 31, 2015 - link
thats a good point. i see this technology being useful in VR goggles that track the users eyes to see where they are lookingWardrop - Sunday, May 31, 2015 - link
I thought the same thing. Seems the idea though is that the pixels on the edges are squished, hence after the distortion, the edges of the screen are actually higher resolution than the centre. With this technique, the resolution at the edges of the screen could be slightly lowered, meaning after the distortion is applied which squeezes together the edges, the end result should be an optically consistent resolution across the entire screen. This is all in theory. It honestly seems more effort than it's worth.twin - Sunday, May 31, 2015 - link
Pretty sure it's good practice to render at a slightly higher res than native to keep the center "blown up" area looking nice post warp. This leads to even more waste on the outer areas. Still, it's probably not worth the effort.MrSpadge - Monday, June 1, 2015 - link
Exactly: they're not looking at making the edges blurry, they're removing information which couldn't be seen there anyway due to the lens warping."This is all in theory. It honestly seems more effort than it's worth."
In many cases people would go to far larger efforts for even a 10% improvement in shader performance. Unless a developer doesn't care about performance and optimization at all.
Dukealicious - Tuesday, June 9, 2015 - link
In a VR headset quite a lot of the screen is not actually visible. For reference the automap, health, shields, and ammo displays in the borderlands games are outside your vision no matter where you point your eyes. Everything is being super-sampled for not just the distortion correction but also since aliasing is very obvious when a screen is magnified by the lenses. The lenses have a sweet spot of clarity and the image also blurs the further you get from it. A FOV of 120 must be set in games to display correctly as well. So the demands of consumer VR with the release of the Vive and Rift is a 2160x1200 supersampled for distortion correction with a 120 FOV, stereo rendered, and a framerate of 90 fps solid to match their 90 hz display so the head tracking doesn't stutter and make you sick.Any performance saving we can take by not rendering high resolutions in the out of sweet spot zones and not rendering at all in the areas invisible to us is much needed in VR.
This type of rendering will make higher FOV headsets in the future worthwhile from a performance stand point.
I will illustrate the current position of performance without this on the Oculus DK2 which is currently available. The DK@ is only 1920 x 1080 and needs to maintain 75 fps for 75 hz. With a GTX 970 I have to turn all the settings in Bioshock Infinite to medium/low for Geometry 3d in the DK2. Not typically a demanding game by any means.
We have all grown accustom to a certain level of graphical fidelity. Having to drop your settings so drastically for VR is sure to disappoint and this Multi-Res Shading is going to mitigate that somewhat.
twin - Sunday, May 31, 2015 - link
Rendering multiple viewports in a single pass is only in dx12, right? Any word on benefits to stereo rendering?