Wednesday, August 1st 2018

NVIDIA Unveils Adaptive Temporal Anti-Aliasing with Ray-Tracing

by
btarunr
 Discuss (71 Comments)
NVIDIA published the first documentation of Adaptive Temporal Anti-Aliasing (ATAA), an evolution of TAA that incorporates real-time ray-tracing, or at least the low light-count method NVIDIA implemented with RTX. Its "adaptive" nature also lets it overcome many of the performance challenges users encounter with TAA in high framerate and rapidly changing 3D scenes, such as in games. Non-gaming scenes, such as those used by real-estate developers, don't face these challenges.

To developers, ATAA promises image quality comparable to 8x supersampling at a cost of under 33 ms frame delay. These numbers were derived on a TITAN V ("Volta"), using Unreal Engine 4. It could take a while for ATAA to make it to games, as developers will need a few months to learn the technique before implementing them in their ongoing or future projects. NVIDIA will introduce ATAA support through driver updates.
Sources: NVIDIA (PDF), Videocardz

Related News

Add your own comment

71 Comments on NVIDIA Unveils Adaptive Temporal Anti-Aliasing with Ray-Tracing

#51
bug
Prima.VeraActually, THE BEST postprocessing AA method outhere, SMAA, adds only 2.5ms delay to a 60fps scene. So you can understand the 33ms horror....
Yes, but doing AA in post means you don't know where the edges are and you're blurring the whole scene. Not sure why you think that method is the best, but ok, at least we got one more number to put those 33ms in some perspective.
Posted on Reply
#52
Prima.Vera
bugYes, but doing AA in post means you don't know where the edges are and you're blurring the whole scene.
This is not how SMAA works...
Posted on Reply
#53
bug
Prima.VeraThis is not how SMAA works...
Sure. Because in post processing you can totally tell an edge from detail on a texture :wtf:
Posted on Reply
#54
goodeedidid
FordGT90ConceptSSAA 4x looks better than ATAA 8x IMO. Why keep reinventing the wheel?


And that's on a Titan V. How bad is it on a 1060 GTX (you know, a card gamers actually use)? Even if it was intended for real estate renders, why use the more expensive ATAA when you can get a better results with SSAA? I think NVIDIA just built another bridge to no where to sell GameWorks.
Probably bad example.
Posted on Reply
#55
mtcn77
Prima.VeraThis is not how SMAA works...
SMAA has separate buffers and antialiasing levels. I like the luma version best. You can use the antialiasing dial, if it is the only active antialiasing filter, just in case standard antialiasing filters aren't present. SMAA works so long as only two borders meet at any pixel, so it doesn't have any extra weights. Tent filters are area weighted, but they are linear. Bilateral is non-linear and weighted, so non-blurring and edge preserving; it is just like rotated grid supersampling at the neighbouring pixels(diagonals versus inlines). If you alternate the sample positions temporally, one could easily make the argument it would make a great checkerboard rendering filter. Further optimisations can be integrated like low resolution shading coupled with coverage sampling.
Posted on Reply
#56
InVasMani
Wait now Nvidia suddenly cares about blurr? For the last 10 years or so nearly every effect optimization they've introduced adds more blur to the image. I swear the thing I miss the most is -15.0 LOD bias from the RivaTuner days back in the GTX280 and earlier era of GPU's. They need a transparent AA technique based on alternate frame rendering that applies a minor unsharp mask to it. Less jaggies overall and a bit sharper on the edges and because of the altnerate frame rendering any ghosting from over sharpening could be minimized as well. Perhaps also a secondary LOD bias based on alternate frame rendering that could be jointly applied thus by splitting it between frames it reduces the shimmering side effect.
Posted on Reply
#57
Prima.Vera
InVasManiWait now Nvidia suddenly cares about blurr? For the last 10 years or so nearly every effect optimization they've introduced adds more blur to the image. I swear the thing I miss the most is -15.0 LOD bias from the RivaTuner days back in the GTX280 and earlier era of GPU's. They need a transparent AA technique based on alternate frame rendering that applies a minor unsharp mask to it. Less jaggies overall and a bit sharper on the edges and because of the altnerate frame rendering any ghosting from over sharpening could be minimized as well. Perhaps also a secondary LOD bias based on alternate frame rendering that could be jointly applied thus by splitting it between frames it reduces the shimmering side effect.
I really don't understand why nVidia doesn't introduce something like TSSAA (Transparancy Super Sampling AA) instead. By far it's the best post processing AA method ever implemented that rivals SSAA with neglijent performance loss.
Posted on Reply
#58
bug
Prima.VeraI really don't understand why nVidia doesn't introduce something like TSSAA (Transparancy Super Sampling AA) instead. By far it's the best post processing AA method ever implemented that rivals SSAA with neglijent performance loss.
I highly doubt anything aware of transparency can be done in post processing stage.
Posted on Reply
#59
Prima.Vera

Sorry, but looking closer to this pic, the technique is crappy to say the least. Just check out the blur. Again. Actually is double the times blurer than FXAA. WTF!?
nVidia loves the blur for some retarded reason....
Posted on Reply
#60
mtcn77
I'm reading up on "Bilateral Rotated Grid Filtering". Somehow it seems harder than it should be. There is an exponential resolve(N^M) hardship. That shouldn't be so difficult. Just look which samples would entitle to 8x grid sampling(off the top of my head, 8 half-pixels apart), should they be normalised. It shouldn't be that hard to select few external samples with weights that approximate a good box filter implementation.
Posted on Reply
#62
mtcn77
Okay, so big thanks to Intel, now I know there are faster ways to filter samples. Namely, moving average(compute shader workload for its less than linear hardship and big kernel utility) and Kawase filter(rotated-grid and non-polynomial).
I liken to think Kawase filter can be combined with some sort of predication bit array(like how VLIW handled multiple streams of code in-flight) that signs a kernel mask for active edge awareness(like QR code). It 'could' approximate a fixed sampling target, for example 8x SSAA per say, if the bit mask is given a LUT which pixel gradients to pass. There can even be an AI algorithm to optimise the shader.
Posted on Reply
#63
Prince Valiant
jabbadapHmh, was that pptx file on the site last time I visited it. Any way it has some more comparing pics of the tech, which does not look that blurry than the pic in the white paper.


There is some GPU times @fullhd for Titan V too
Thanks for those. Seems like there are times where x8 can look alright but the time penalty is rough.
Posted on Reply
#64
bug
Prince ValiantThanks for those. Seems like there are times where x8 can look alright but the time penalty is rough.
I would love to see great AA having zero performance impact, but HQ AA still has a saving grace. You really only notice the quality of AA in titles that play well at 30fps. So if you're playing an adventure game or something, you can still up the AA quality and enjoy the game. Ymmv, of course, but as a rule of thumb, I think it holds up pretty well.
Posted on Reply
#65
xorbe
I wish demo pics would include a theoretically proper/perfect AA rendering.
Posted on Reply
#67
Prima.Vera
bugSure. Because in post processing you can totally tell an edge from detail on a texture :wtf:
Dude, AA it's not used to smooth the textures edged, lol (unless you are nVidia :D) . You're mixing poligons, texturing and post processing....
Posted on Reply
#68
StrayKAT
What games use ATAA? I like TAA from what I've seen.
Posted on Reply
#69
jabbadap
StrayKATWhat games use ATAA? I like TAA from what I've seen.
None, it's on one of the publications of nvidia research team. And they made technical paper public just August 10 2018 on High-Performance Graphics conference at Vancouver. But if Quadro RTX 6000 can do ATAA x8 5.6ms at FullHD(down from TitanV:s 18.5ms), that really does not sound too bad or heavy. Well yeah, still they should show some real world demo video with it, before making any conclusion.
Posted on Reply
#70
mtcn77
mtcn77Okay, so big thanks to Intel, now I know there are faster ways to filter samples. Namely, moving average(compute shader workload for its less than linear hardship and big kernel utility) and Kawase filter(rotated-grid and non-polynomial).
I liken to think Kawase filter can be combined with some sort of predication bit array(like how VLIW handled multiple streams of code in-flight) that signs a kernel mask for active edge awareness(like QR code). It 'could' approximate a fixed sampling target, for example 8x SSAA per say, if the bit mask is given a LUT which pixel gradients to pass. There can even be an AI algorithm to optimise the shader.
More to the point: the bit mask I cited was referencing Coverage Based Antialiasing filter found in "CoD:IW". While it is possible to leave it at the bit mask stage for edge-awareness across neighbouring pixels, I intend to generalise this like a spatial domain filter, but as a 'moving average' that differentiates objects by the object ID "UAV" ports. Such a formulation is easily contrivable. You just average-move across rows until the 'Object ID' tells you that coverage is not continuous per UAV's. You then tell the renderer to lessen denominator weight by 1 per each sample weighted until the edge pixel that has just n+1 weights rather than 2n+1 that cover pixels(pixels>>2n samples).
If you have moving average, you can have a gaussian which you can build a fast bilateral filter with.
Posted on Reply
Add your own comment
Nov 27th, 2024 12:48 EST change timezone

Latest GPU Drivers

New Forum Posts

Popular Reviews

Controversial News Posts