Nvidia developing Subpixel Reconstruction Anti-Aliasing (SRAA) to combat ATI's MLAA

[taltamir]

Check out Dead Space or Borderlands with MLAA and compare it with MSAA or SSAA. If anyone thinks the usual AA modes look superior in those two games, they must be blind.

SSAA is not a "usual AA mode", it has been removed with DX8 and only recently made a limited comeback. MSAA is the now obsolete predecessor to CSAA, a far more efficient algorithm, which is now often combined with TrAA.

here is an actual comparison of MLAA on the games you mentioned:
http://translate.google.com.ni/tran...a-bastante-a-los-juegos-de-pc-gracias-al-ps3/
I don't like the way they handle their data (only provide FPS data for one game, starcraft. From another website, and compared to MSAA 4x and SSAA of all things).
But they have somewhat useful pictures. (although they only compare MLAA to no AA at all, except for deadspace where their comparison image is zoomed in way too much. Which masks the negative effect of blurring while intensifying the effects of jaggedness).

You can see at the bottom that in deadspace specifically, MSAA 4x produces a sharper, but more jagged image then MLAA. MLAA does reduce jaggedness more, but at the cost of blurring the image some. You can combine MSAA 4x and MLAA in that game to get even less jaggedness (although there is still slight blurring).

I would be curious to see how it compares to 8x MSAA, 16x MSAA, CSAA, or SSAA. Does using MLAA + 4x MSAA cost more (in terms of GPU performance) than just using 8x MSAA or some form of CSAA? And also, in what games. Some games might be exceptionally suited for one type of technology over another, it happens.

And if given a choice, I would just use the driver option of replacing all MSAA with CSAA (the two are very similar and easily interchangeable, but there is a significant difference which makes CSAA a lot better and very cheap to implement. check my AA explanation for details).

All that being said, on that specific game MLAA actually manages to look less jagged than 4x MSAA. I did say it looks worse then any other form of AA but it is subjective if the sharpness of 4x MSAA is worth the extra jaggedness; and it might be ever more pronounced with 2x MSAA. So perhaps on some games MLAA is better for some tastes than low levels of MSAA.

Another thing is, I dismissively said "there is no need to answer MLAA" and I don't know where that came from. Maybe because its so late. Such a statement completely disregards the actual purpose of it. To provide AA for games who otherwise have no AA capability whatsoever. MLAA/SRAA provide a fallback solution in those cases. Theoretically it should even be applicable to 2d sprite games like PS:Torment. I really hope that is the case, and can't wait to test that out.

 

[thilanliyan]

I tried MLAA for a bit and it looks okay, other than for text.

 

[busydude]

I tried MLAA for a bit and it looks okay, other than for text.

For the games I played, it works pretty well for Borderlands, GTA IV and Darksiders. Yeah.. if a game has small text.. its going to be an issue.

 

[BFG10K]

I read through the nVidia statement, SRAA is exactly the same as MLAA in all the details they provided.

No. SRAA uses sub-pixel data while MLAA does not. I’m not saying SRAA will be better than traditional AA methods, but it should be better than MLAA.

And if given a choice, I would just use the driver option of replacing all MSAA with CSAA (the two are very similar and easily interchangeable, but there is a significant difference which makes CSAA a lot better and very cheap to implement. check my AA explanation for details).

You can’t replace MSAA with CSAA. The two are not interchangeable because you need both. CSAA by itself couldn’t anti-alias anything because it has no color data.

 

[Mistwalker]

Is there any hardware-related reason why Nvidia cards couldn't run MLAA or AMD cards SRAA?

 

[BFG10K]

Is there any hardware-related reason why Nvidia cards couldn't run MLAA or AMD cards SRAA?

Not that I know of.

 

[taltamir]

No. SRAA uses sub-pixel data while MLAA does not. I’m not saying SRAA will be better than traditional AA methods, but it should be better than MLAA.

I see now, it collects subpixel visibility data during rendering without modifying rendering to facilitate better accuracy when it applies the post processing.

Would you say that that would make it less flexible and easy to implement compared to MLAA? While providing better IQ?

 

[SirPauly]

I see now, it collects subpixel visibility data during rendering without modifying rendering to facilitate better accuracy when it applies the post processing.

Would you say that that would make it less flexible and easy to implement compared to MLAA? While providing better IQ?

Considering it is collecting subpixel data, probably shine more while moving, when compared to MLAA. The key though will nVidia be able to allow end-users to simply enable it or is this for developers to use for their engines or titles? Don't know, considering the lack of data.