AMD Morphological AA Performance and IQ Review @ [H]

FrgMstr

Just Plain Mean
Staff member
Joined
May 18, 1997
Messages
55,620
AMD Morphological AA Performance and IQ Review - We take an in-depth look at AMD’s new Morphological AA technology introduced with the Radeon HD 6870 and Radeon HD 6850. We will evaluate performance and image quality with the new antialiasing feature. This new feature is poised to provide an improved gameplay experience.
 
I gotta say I love the idea of having an AA "nuclear" option, for games in which is just isn't supported.
 
Great job.

Does MLAA use less VRAM than your typical MSAA? (I kind of assume so based on how it works as a filter after the rendering). VRAM seems to be the limiting factor in eyefinity scenarios where AA is turned on. I know the performance hit was pretty high at high resolution, but I'm wondering if this will be the future go to AA method for high resolutions in xfire setups.
 
Last edited:
MLAA may not blur the textures but it still makes everything look more blurry than anti-aliased to me when going by the screenshots.

I'd go with 2x MSAA or No AA before I'd use MLAA. The image quality of MLAA looks horrible.
 
MLAA may not blur the textures but it still makes everything look more blurry than anti-aliased to me when going by the screenshots.

I'd go with 2x MSAA or No AA before I'd use MLAA. The image quality of MLAA looks horrible.

I second this. I turned it on and loaded up WoW, thought I messed something up because everything looked super blurry. I like the idea of the feature, and haven't checked it out in an FPS game yet, but from what I've seen I didn't like it.
 
Great review! The IQ comparisons were extremely helpful, and while MSAA and MLAA do trade blows, I think the ability of MLAA to hit all games as well as everything in the scene is going to be a boon for gamers. And as you guys said, it's optional. I think those complaining about it looking "blurry" are either reading too much into screenshots or are just so used to hard pixellated edges in games that to finally have a fully anti-aliased scene feels unnatural to them.

I'm really looking forward to trying this out on games like Mass Effect that have lots of jaggies and issues with adaptive AA.

I know they've said that this feature will be backported to the 58xx series, but any word on roughly when they're planning to do so?
 
I second this. I turned it on and loaded up WoW, thought I messed something up because everything looked super blurry. I like the idea of the feature, and haven't checked it out in an FPS game yet, but from what I've seen I didn't like it.

I don't see why you would use MLAA in WoW. I would use it in a game like Mass Effect 1/2.
 
I second this. I turned it on and loaded up WoW, thought I messed something up because everything looked super blurry. I like the idea of the feature, and haven't checked it out in an FPS game yet, but from what I've seen I didn't like it.
I thought WoW looked blurry with traditional Supersampling, personally. I don't know if that's an MLAA problem or a problem with the game engine, to be honest.

I play WoW at 8X AD MSAA, because Supersampling looks worse.
 
Disappointed that this hits such a wall at high resolutions. Basically from what I'm hearing is that for Eyefinity, it's useless. I tried it with the hacked 5870 drivers on my crossfired setup and it was single digit frames. I was hoping it was just the unsupported drivers but from the article it seems as though it's just the huge size that this filter can't process.

Would be great for the single-screeners, though.
 
Morph AA sounds fantastic. Finally a way to force AA on games that don't support AA. Oh and I guess for AMD this is also a good thing because now their video cards can still do some form of antialiasing even if Nvidia vendor-locks them out of a game's normal AA capability. It would be cool to see how and if Morph AA works and looks in Batman:AA because that game was notorious for that very reason!
 
I never liked AA and always disable it, the high contrast look is well worth the higher frame rate. At some point in the future image enhancing features like this will come into play but not till they can make it work much faster. MLAA is a great thing but it is still too slow IMHO :(
 
I think those complaining about it looking "blurry" are either reading too much into screenshots or are just so used to hard pixellated edges in games that to finally have a fully anti-aliased scene feels unnatural to them.
No, that's not it... You really can't see a difference between this and supersampling?

Supersampling adds information to a frame; morphological AA subtracts it. It may get rid of aliasing artifacts, but in terms of image quality, it's a step sideways at best. It'd take some pretty horrific shader aliasing before I'd favour it over 2xMSAA...
 
i'm not willing to sacrifice image quality for better edges.
 
Not bad, First I want to say that I really enjoyed the article. It was very well put together and easy to follow.

I like everything about MLAA and applaud AMD for bringing it to their customers. I can see this being huge in eyefinity when you have a crossfire set of 6950s or 6970s rendering across 3 16x8 or 19x10 displays. On the 6800 cards it seems they simply hit a wall at the higher resoolutions which is understandable, those gpu's are for 19x12 and below, although they are performance beasts in that class.

I can only assume nvidia will develop something similar soon. Kudos to AMD again for a very great way to give more visual quality and very low performance impact at the same time.
 
Interesting read, thank you.

I'll be able to test it when 10.11 hits I suppose. AMD are not including Vista .infs with their current crop of hotfix drivers. Yes, that includes 10.10d.
 
Great article Brent and Kyle. Eagerly anticipating this one for quite a few days now :D

"The only improvement I can see is that game developers could use this technology inherently in their games; in the GPU pipeline process, so that the AMD drivers wouldn’t have to apply it on post. This could speed it up a great deal and really allow game developers to control the intensity and color and sharpness on edges. "

This AMD, print it out and post everywhere the "MLAA team" works. :)


"This is a welcomed new optional technology, and we look forward to exploring its performance on the Radeon HD 6900 series."

I very much agree with this. MLAA is very promising.


Any elaborations on potential experiences of input lag on the horizon? I guess there was none in this particular game. But seeing there is a 10.10d out, I thought maybe this could be an important thing to discuss in future and further testing?

MLAA may not blur the textures but it still makes everything look more blurry than anti-aliased to me when going by the screenshots.

I'd go with 2x MSAA or No AA before I'd use MLAA. The image quality of MLAA looks horrible.

Horrible is a bit far IMHO. The blurryness and lost definition has me concerned as well. It is there and it is noticeable for sure. That being said, one thing is pixel peeping stills another entirely is moving images as we see them when we play games. I agree that MLAA does not compare to SSAA or even MSAA in most cases.

The way I will be using this, is seeing how it responds to the game in need, with the consideration of the existing textures of the game. If these are A: not that great, or B: unsharp-masked to death I would very much welcome some MLAA, even losing some of those tight and highly specular surfaces. An unnaturally sharpened and high contrast picture is as much an eye sore for me than those "void textures" with fuzzed out blurry non detail, that pops up in a lot of (console)games.

Like everything else, its a matter of judging what is best for a given situation. MLAA just gives the end user more options to tailor to his needs. Its a win win.


Disappointed that this hits such a wall at high resolutions. Basically from what I'm hearing is that for Eyefinity, it's useless. I tried it with the hacked 5870 drivers on my crossfired setup and it was single digit frames. I was hoping it was just the unsupported drivers but from the article it seems as though it's just the huge size that this filter can't process.

Would be great for the single-screeners, though.

I am fairly confident that AMD has the 2560 crowd in mind for the 69xx series. I understand your worry about Eyefinity though. Will be interesting to see if MLAA is possible for 3 x 1920 with this generation of cards, or if its a 28nm thing.
 
Last edited:
I wonder if this is something that can be produced by all those unused cores on our CPUs???
 
I wonder if this is something that can be produced by all those unused cores on our CPUs???

The code for this is based in HLSL for DirectCompute.

I asked this question and this is Terry's response about that:

The answer is that it initially stared as an HLSL shader and uses Compute functionality in the chip such as the LDS, however what is stored in the driver is native to our architecture and can be ported across API’s and platforms.

I don't think we'll see it accelerated via CPU cores, it just doesn't make sense, this is a GPU accelerated feature utilizing technology in the GPU (Local Data Stores.)
 
I personally do not like MLAA. I don't like the fact that it applies to everything displayed on the screen. Text is blurry, huds are blurry, ect...
 
I thought WoW looked blurry with traditional Supersampling, personally.
SSAA blurs textures unless you drop the mipmap LOD bias. ATI neglected to do this in the original drivers for the 5000 series, but that might have changed by now; in any case, you should be able to force it with ATI Tray Tools or something similar.
 
Now the question is, when will ATI officially (I know we have hacked drivers) implement it for 5xxx?

I'm guessing they won't for awhile to try to stimulate sales for 6xxx?
 
No, that's not it... You really can't see a difference between this and supersampling?

Supersampling adds information to a frame; morphological AA subtracts it. It may get rid of aliasing artifacts, but in terms of image quality, it's a step sideways at best. It'd take some pretty horrific shader aliasing before I'd favour it over 2xMSAA...

That's incorrect, supersampling does do some blurring and has some loss of information, particularly on fine mesh patterns etc, it's just the nature of the beast. I wanted to see some screen comparisons of MSAA with MLAA compared to SSAA, but a good review.

I did see some blurring with MLAA and loss of information (check fences), even a little more than SSAA but it will depend on the game whether this is useful or not. That's also something that could be tuned by changing the shader algorithm. Racing games traditionally have tons of aliasing, so this was a perfect game to test it out on.
 
That's incorrect, supersampling does do some blurring and has some loss of information, particularly on fine mesh patterns etc, it's just the nature of the beast. I wanted to see some screen comparisons of MSAA with MLAA compared to SSAA, but a good review.

I did see some blurring with MLAA and loss of information (check fences), even a little more than SSAA but it will depend on the game whether this is useful or not. That's also something that could be tuned by changing the shader algorithm. Racing games traditionally have tons of aliasing, so this was a perfect game to test it out on.

Yes, I find racing games benefit greatly from AA methods, especially looking out into the distances on the tracks. I'm a huge racing game fan.

Anyway, the way I see it is that MLAA is a less expensive alternative to SSAA, as we have shown, it isn't as good as 8X AD SSAA, but its playable, where 8X AD SSAA isn't. It is more comparable to 2X AD SSAA, and in some cases 2X AD SSAA will look better, and in some cases (especially in the far distance) MLAA may look better, or at least the same. The catch is, MLAA is also faster than 2X AD SSAA. So that makes it more usable.
 
No, that's not it... You really can't see a difference between this and supersampling?

Supersampling adds information to a frame; morphological AA subtracts it. It may get rid of aliasing artifacts, but in terms of image quality, it's a step sideways at best. It'd take some pretty horrific shader aliasing before I'd favour it over 2xMSAA...

Well of course I would love to run 8x SSAA on everything I play, but not only is that impractical from a performance standpoint, it also just flat out doesn't work. Lots of games have various weird issues with MSAA, SSAA, adaptive AA, or all three. Mass Effect 1/2 are perfect examples of this. In cases like that, MLAA "photoshopping" the final image would actually be preferable.

As for being able to tell the difference - like I said, both traditional AA and MLAA have wins and losses in the comparison shots shown in this review. I can't personally vouch for MLAA in the real world yet because I don't own a 6 series card.

If someone wants to point me to these hacked 5 series drivers I'll be happy to weigh in :D
 
Be warned, user feedback is that the hacked 5xxx driver introduces input lag. I recommend waiting for the official driver.

Someone in this thread I think asked about that in my testing, I did not experience any input lag with either 6800 series card with 10.10c driver.
 
Can you elaborate on this?

The fuzziness bothers me, AA should make.straight lines but MLAA looks worst then MSAA because the image quality goes down. I tried all the AA's but they all still have those flaws, some games it works while others it doesn't. Depending on the texture quality I can't tell the difference unless the same has poor resolutions cinematics in like Star Craft2. Even after the patches and new drivers, I still have squarish bullets.
 
Be warned, user feedback is that the hacked 5xxx driver introduces input lag. I recommend waiting for the official driver.

Thanks Brent. Any chance of them screwing up my install or causing hardware damage? I'd only be using it out of curiosity and to get a first-hand look at MLAA. Not actually planning to spend any serious play time with them.
 
The fuzziness bothers me, AA should make.straight lines but MLAA looks worst then MSAA because the image quality goes down. I tried all the AA's but they all still have those flaws, some games it works while others it doesn't. Depending on the texture quality I can't tell the difference unless the same has poor resolutions cinematics in like Star Craft2. Even after the patches and new drivers, I still have squarish bullets.

I'm not saying it is perfect for every situation, but I actually find MLAA looking better than 8X MSAA on polygon edges, in my zoomed up shot on the IQ page you can see how MLAA is blending better than 8X MSAA. It also looks better in normal view. I'll give you the fact that it doesn't look better on every object, case in point the tree. But I wouldn't make the case that it just makes everything look fuzzy and reduces image quality. IMO, it reduces aliasing, edge, alpha and shader, thus improving IQ.
 
I'm not saying it is perfect for every situation, but I actually find MLAA looking better than 8X MSAA on polygon edges, in my zoomed up shot on the IQ page you can see how MLAA is blending better than 8X MSAA. I'll give you the fact that it doesn't look better on every edge, case in point the tree. But I wouldn't make the case that it just makes everything look fuzzy and reduces image quality. IMO, it reduces aliasing, edge, alpha and shader, thus improving IQ.

how about this, they mix both. MLAA for long distance images and MSAA for close range. MLAA post processing and MSAA is pre processing. you can have your cake and pie the same time.
 
how about this, they mix both. MLAA for long distance images and MSAA for close range. you can have your cake and pie the same time.

The only way to do that would be for MLAA to be part of the rendering process in the pipeline, and not a full-scene post-process effect. Unless they could tweak the filter so that it somehow detects edges more in the center of the image, but don't ask me if that's possible.

If MLAA was implemented by the game developer, then what you say could easily be done, and I agree, it would look phenomenal.

The current method right now supported by AMD is a nice option for games that don't support AA at all, or have some kind of funky AA that doesn't work well. The technology has the benefit of working in every game no matter what.
 
the Russian next to me said, "maybe", at least he isn't going on a rant anymore.
 
Starting with the Voodoo 3 and beyond, 3dfx used post processing to enhance its color precision from 16-bits to 22-bits, and with the V5, which was a full 32-bit RGB, 3d-output card, 3dfx used post processing for FSAA and for its T-buffer effects (T-buffer required the V5 because the V5 came standard with two GPUs running in tandem.) Since all of these things were accomplished by post processing, they would not work with standard screen grabbers which only grabbed the frame-buffer contents. Like with MLAA, you had to have a special screen-grab utility which was capable of grabbing the entire post-processed frame as opposed to merely what was in the frame buffer. (Although, from what I understand the MLAA screen grabs here are externally run through a program which attempts to duplicate the on-screen effects of MLAA post processing.)

I remember Anand did an absolutely poor-quality review of the V3 because he did not use the required screen-grab software and published several V3 screen shots which looked, in a world, horrible, with far poorer image quality than the V2. Anand did notice the difference and said something very lame, like, "I did notice, however, that my screen shots looked nothing like what I saw on the screen." Pity he didn't think to check that out with 3dfx at the time prior to publishing his review...;)

http://www.itnetcentral.com/tech/voodoo3-3000-326.html

When grabbing screenshots, keep in mind 3dfx has a unique way of converting data in the frame buffer before we send it to the screen to be displayed. As a result, we store "unfiltered" data in the frame buffer - and the filtering is performed at scan-out on the way to display. This improves the quality of the picture displayed on the monitor, but doesn't actually change the data in the frame buffer. Typically when you are doing a frame grab, you are just grabbing data directly from the frame buffer, which is "unfiltered" so the quality of raw images grabbed from the frame buffer is actually lower than what is actually displayed. HyperSnap, a frame grabbing utility, actually implements our filtering technique on their frame grabs, producing a more accurate representation of what the user would see on their display by actually running our filter on the data and writing out the filtered, instead of the "raw" data. Get HyperSnap at http://www.hyperionics.com/

Wonder if something like Hypersnap might exist to properly grab the on-screen MLAA images?
 
I found myself going back and forth in that review. It would have been nice to see that page of screenshots of what MLAA does BEFORE worrying about which of the five pages of graphs I should care about. :)

So, if (subjectively) I felt MLAA sits somewhere between 4X and 8X MSAA, I could safely ignore 2X MSAA... although 2X MSAA with MLAA would still have been nice to see.

Regardless, I do like the option of being able to turn on a global setting without having to worry overmuch about game-specific implementation. Thanks for the comparo!
 
MLAA reminds me of how Ubisoft added "Edge Smoothing" to Ghost Recon Advance Warfighter a few years ago because AA wasn't supported. It smoothed the entire image including the HUD. It really helped spot enemies from a distance but also made the compass, minimap, etc difficult to read at high resolutions because these items were small and then sorta mangled. This was the only thing I wish the article had gone into detail more... although you can kind of see how MLAA is affecting the digital speed readout on the steering wheel in this pic:

1288596806JDsZFK1bAE_9_9_l.png


Hard to describe it, but the numbers have kind of a "caved in/mangled" look to them.

Also it would be nice to see the impact on performance for eyefinity resolutions, like 5040x1050, rather than the 30" resolution. Otherwise, nice article.
 
Interesting feature but I don't think I'd use it on anything but older games or games that don't support AA (either in-game or by forced).
 
That's incorrect, supersampling does do some blurring and has some loss of information, particularly on fine mesh patterns etc, it's just the nature of the beast.
Not as I understand it. You're using more information about the underlying scene (i.e. taking more samples) to build your frame, so you should always get a more accurate picture as a result, assuming that your additional samples actually do contain additional information.

This is always the case with geometry, and AFAIK, always the case with per-pixel shader effects. The only place it breaks down is the textures - your additional samples may all hit the same texels in a low-res, pre-antialiased mipmap, so you wind up with the same information as your non-AA frame, only smeared around a bit.

But the solution is simple: take your samples from the higher quality mipmaps. Drop your mip bias, enough to give you texture aliasing artifacts in a non-AA frame, and let the SSAA antialias them for you.
 
Back
Top