It seems that you're using an outdated browser. Some things may not work as they should (or don't work at all).
We suggest you upgrade newer and better browser like: Chrome, Firefox, Internet Explorer or Opera

×
avatar
Lehti: Ok, I think I got it. Is there a more common term for the algorithm other than "übersampling"?
avatar
cjrgreen: It used to be called "full-screen anti-aliasing" (FSAA). It was dropped by GPU makers because it is so inefficient, then reinstated in recent GPUs because they have enough computing power to do it (Fermi and Northern Islands GPUs have it).

It's pretty inefficient to compute, because you need to render every frame 2 to 4 times, rather than traditional AA, which applies a more capable filter only to the edges in the frame designated for AA. Also, its performance is that of a simple first-order moving-average filter, which does little except soften the image's focus.
No, as mentioned earlier in the thread FSAA renders the scene ONCE at a very high resolution then scales it down. According to CDPR's description of ubersampling it renders the same resolution frame multiple times and then composites them somehow... the question is how, since there must be some jitter or something to ensure that the frame being rendered multiple times is not identical and thus not helpful.
avatar
cjrgreen: It used to be called "full-screen anti-aliasing" (FSAA). It was dropped by GPU makers because it is so inefficient, then reinstated in recent GPUs because they have enough computing power to do it (Fermi and Northern Islands GPUs have it).

It's pretty inefficient to compute, because you need to render every frame 2 to 4 times, rather than traditional AA, which applies a more capable filter only to the edges in the frame designated for AA. Also, its performance is that of a simple first-order moving-average filter, which does little except soften the image's focus.
avatar
Vash63: No, as mentioned earlier in the thread FSAA renders the scene ONCE at a very high resolution then scales it down. According to CDPR's description of ubersampling it renders the same resolution frame multiple times and then composites them somehow... the question is how, since there must be some jitter or something to ensure that the frame being rendered multiple times is not identical and thus not helpful.
No, that is different; what you are calling FSAA is actually supersampling. Supersampling would permit more complex filtering as the image is downscaled; for example, a sinc or Lanczos filter.

FSAA as originally described renders each pixel 2, 4, or more times at offsets of +/-1/N pixel. It does not render the image at a different scale. The only difference between FSAA and the way Ubersampling has been described is that Ubersampling renders the entire frame before filtering, rather than working pixel by pixel. This may be computationally more efficient on massively parallel GPUs at the expense of requiring additional memory for the complete frame.

FSAA reference, with comparisons to other AA methods: http://homepage.mac.com/arekkusu/bugs/invariance/FSAA.html

Either way you compute it, it's a first-order moving-average filter, nothing more, and there are far better ways to do antialiasing.
Post edited May 19, 2011 by cjrgreen
avatar
Vash63: No, as mentioned earlier in the thread FSAA renders the scene ONCE at a very high resolution then scales it down. According to CDPR's description of ubersampling it renders the same resolution frame multiple times and then composites them somehow... the question is how, since there must be some jitter or something to ensure that the frame being rendered multiple times is not identical and thus not helpful.
avatar
cjrgreen: No, that is different; what you are calling FSAA is actually supersampling. Supersampling would permit more complex filtering as the image is downscaled; for example, a sinc or Lanczos filter.

FSAA as originally described renders each pixel 2, 4, or more times at offsets of +/-1/N pixel. It does not render the image at a different scale. The only difference between FSAA and the way Ubersampling has been described is that Ubersampling renders the entire frame before filtering, rather than working pixel by pixel. This may be computationally more efficient on massively parallel GPUs at the expense of requiring additional memory for the complete frame.

FSAA reference, with comparisons to other AA methods: http://homepage.mac.com/arekkusu/bugs/invariance/FSAA.html

Either way you compute it, it's a first-order moving-average filter, nothing more, and there are far better ways to do antialiasing.
"Super Sampling anti-aliasing (SSAA),[2] also called full-scene anti-aliasing (FSAA),[3"
http://en.wikipedia.org/wiki/FSAA#Full-scene_anti-aliasing

And yes, the link is sourced. You can find like 20 other references also if you try Googling it yourself.
avatar
Fizzl: It's the setting everyone who doesn't have multiple graphics cards should turn off!

I couldn't see the difference on 1920x1080 so I assume its meant for large screens/eyefinity.
I have a single card and Ubersampling works fine.
avatar
Fizzl: It's the setting everyone who doesn't have multiple graphics cards should turn off!

I couldn't see the difference on 1920x1080 so I assume its meant for large screens/eyefinity.
avatar
mojoman69: I have a single card and Ubersampling works fine.
Could you give me your specs? I just wanna compare specs. Thank you.
There is no single GPU that can run The Witcher 2 maxed with Ubersampling On at a good framerate (30fps or more). Not even the GTX 680 can do it.

You need GTX 470/560TI in SLI minimum and I can definitely tell the difference on my single GTX 470.

Everything looks much smoother but I'm wondering if AA couldn't have been implemented in a more traditional way with small steps like 2x, 4x, 8x, 16x etc.

It seems to me that Ubersampling is applying a lot of AA which kills the GPU even though it's not necessary.
avatar
tranceland: There is no single GPU that can run The Witcher 2 maxed with Ubersampling On at a good framerate (30fps or more). Not even the GTX 680 can do it.

You need GTX 470/560TI in SLI minimum and I can definitely tell the difference on my single GTX 470.

Everything looks much smoother but I'm wondering if AA couldn't have been implemented in a more traditional way with small steps like 2x, 4x, 8x, 16x etc.

It seems to me that Ubersampling is applying a lot of AA which kills the GPU even though it's not necessary.
Ubersampling is a "dumb" brute-force form of AA. Effectively, it doubles the GPU's workload. The game has other less taxing AA settings you can use instead. But don't force AA through your GPU's control panel; the game does not get along with forced AA.
avatar
tranceland: There is no single GPU that can run The Witcher 2 maxed with Ubersampling On at a good framerate (30fps or more). Not even the GTX 680 can do it.

You need GTX 470/560TI in SLI minimum and I can definitely tell the difference on my single GTX 470.

Everything looks much smoother but I'm wondering if AA couldn't have been implemented in a more traditional way with small steps like 2x, 4x, 8x, 16x etc.

It seems to me that Ubersampling is applying a lot of AA which kills the GPU even though it's not necessary.
I got a solid 30 fps with my GTX 680 with the basic clocks. There is probably some part of the game which that can drop i presume, but most of the time it's 30.

I suppose "ubsersampling AA" use the Full Scene Anti-Aliasing which calculate a higher resolution of rendering & then adapt this high resolution rendering at your monitor's resolution.

That's why aliasing disappear & everything is sharp & soft.

Too bad we can't specify the MSAA (2x, 4x, 8x, etc..) & FSAA lvl from 100% to 200% (where 100% is your standard monitor resolution).
ArmA 2 got all these options & it's pretty cool to adjust the rendering to the GPU capabilities.

But with DX11 next generation game (Witcher 3), all this gonna disappear.
Post edited May 08, 2012 by Paul-Hewson
I get 30 with an overclocked 670gtx on uber. Switch it off its a solid 60fps.
I was finally able to run the game at playable FPS (30+ FPS at all times) with everything maxed and ubersampling enabled at 2560x1600.

All it took was GTX680 4GB SLI. :-/