Radeon HD5xxx filtering issue

[Xarick]

if it is a hardware limitation could they give the option to disable angle independence of their af to make it go away?

 

[Blastman]

Or:
ati: http://h-4.abload.de/img/ut3-amdjk0rr.jpg
nv: http://h-3.abload.de/img/ut3-nvidiajt33y.jpg

To me, it pretty much seems like AF+bilinear instead of AF+trilinear (judging from the ati clips) or some kind of bug. It does not depend on whether you force AF via driver or application!

If the ATI card was actually doing bilinear on the whole screen, then the bilinear texture band would extend across the whole screen -- but it doesn't. The band is only evident on the one floor grate. The NV card is having problems rendering the top portion of the grate properly too, which suggests some kind of texture problem in the game.

On the ATI card, the top half of the floor grate looks like a completely different texture as compared to the bottom half. Since in trilinear filtering you're trying to blend (interpolate) the transition between the mipmaps, my thinking is that if a different texture (the wrong one or an improperly coded one) is being called, then the ATI driver would be unable to interpolate properly between them and you would end up with what looks like a bilinear transition between the two different textures. This would simply be because there are two completely different textures on either side of the mipmap and they can't be blended properly. This is not a lousy AF/trilinear issue. In any case, I think it's pretty clear that something is wrong with the textures on that grate either in the game or how ATI's drivers are handling it -- not with ATI's general AF and trilinear IQ as the rest of the screen looks fine. But to try and use this rendering error on a small portion of the screen in making the case that ATI's AF or trilinear filtering is generally worse than NV's is ridiculous.

Here are the comparison shots, from TrackMania, the game I made the videos from.
ati: http://www.abload.de/img/tmu-amdjy2ye.jpg
nv: http://www.abload.de/img/tmu-nvidiajk2db.jpg
Nvidia, no rings, ATI visible rings!

The last picture in that sequence definitely appears to be bilinear filtering (harsh mipmap transitions). There is no way any of ATI’s AI “optimizations” would be degrading the filtering to straight bilinear unless something is broken in the program or drivers. Bilinear with high levels of AF sticks out like a sore thumb. Again, to try and use this as an example of ATI's lousy AF or trilinear filtering (a case where something is broken or not working properly) compared to NV simply doesn't fly.

I hope you're finally convinced that we're really facing a problem here! The more of you stop ignoring (yeah yeah, we all know it only occurs in extreme situations, so you can leave that out now) this, the more likely we will see this fixed in hd6xxx series.

If these bilinear transitions were a common problem and not something broken in a few rare instances, you could be sure the gaming sites would be all over this like a dung beetle on a fresh pile of poop as a major IQ problem (as I’m sure Nvidia would let them know about it). Given that you don’t, this appears more like a couple of isolated cases where something isn’t working properly. This wouldn’t be the first time something is broken or not working properly in a game. There was a recent case of someone complaining about texture shimmering on their GTX 480 in the game Singularity …

techspot review of Singularity Also, I've had some nasty texture streaming issues with this game (on my GTX 480).

(Techspot) … Upon first run we found it odd that the graphics on Singularity looked very low-res. Something was definitely wrong and because we were able to test the game before its official U.S. release we suspected it could be related to a driver bug. Eventually we learned that there is a bug with texture loading that is present in the shipping version of the game.

Once again, my intention was not to call the ATI filters crap, but to point out that they are clearly inferior to Nvidia's filters on current hardware and no review bothers mentioning it.

I think you’re going to have to make a much better case than an isolated instance (or 2) of lousy filtering on a couple of grates and Trackmania (where something is evidently "broken" and thus is not representative of ATI's filtering quality) to make the case that ATI’s filtering is generally inferior to NV’s.

Consider another POV …

nvidia.com

I´m using an GTX 480 with the 197.75 driver.

While playing warhammer online there are two major flickering issues.

First:

Texture-Flickering on Groundtextures. There are several modi in the nvidia controlpanel but non of them could fix the problem. Its completely unimportant which AA oder AF you are using. You will still get an angle dependent flickering. Only unsupported mods like 4xS can solve it.
But with the compromise of loss of vertical structurefiltering und most important gpu-heat and noise.

Second:

Clipping Failures:

So, as written in other threads i read, thats seems to be a nvidia-only problem. In short: Structures beneath upper elements are shinig through with the result that you get an intense flickering: Example: Chosen armor plus cloak

I just tried an ATI-HD5870 and I must say: At the moment the filtering is much better and more exact than the nvidia one.

In terms of performance i will stick with nvidia hoping that this broadband-flickering problem will be fixed, cause it seems that may games are suffering from this issue.

Best regards

If the IQ issue was taken up in earnest (by the hardware sites), and we really started to scrutinize both sides rather than the few instances where something is broken in a game and/or not working properly (whether ATI's fault or the game developers fault), and that is not indicative of ATI's general filtering quality, you may not find that NV would come out ahead overall.

 

[Aristotelian]

Off-topic:
In case you're wondering. Someone with a bit knowledge in German language was not happy with my previous nick (too scatological). I can fully understand that someone could have easily misinterpreted it.

There was nothing to misinterpret. So no.

 

[KARpott]

I think you’re going to have to make a much better case than an isolated instance (or 2) of lousy filtering on a couple of grates and Trackmania (where something is evidently "broken" and thus is not representative of ATI's filtering quality) to make the case that ATI’s filtering is generally inferior to NV’s.

Games in which I observed the effect: TrackMania, Ut3, MirrorsEdge, Doom3, Prey, MaxPayne2, GTA4, Crysis, EliteForce2, HalfLife2, and I think there are many more, so I count at least 10 games, 8 more or less different engines and 3 APIs. I can show (and have shown) you screenshots if you like, many can be found in that 3DCenter-Thread I have linked to. Do you think they all have the same problem with the driver, on totally different engines and APIs, and why is it that it should only happen with games I have bought? Did you even follow the whole discussion here? Even with simple AFTesters, I could reproduce the banding (as I mentioned here in this thread!!!). Rare instances? Maybe, but not rare enough to believe that all the game developers did something wrong (apart from choosing bad content). A driver issue? Maybe, I never ruled that out, but unlikely imo. And that different texture thingy. Do you even believe this yourself? Why should it load a different texture with 16xAF when it loads the right texture with 2xAF and only "partly" loads the right texture with 4xAF? (ah yes, you probably didn't follow the discussion and thus don't know that it's not happening with 2xAF). I'm not the expert here, but from what I could see with my limited shader experience in source, something that strange wouldn't happen.
Even your quotes of my posts show that I never used these pics as an example of ATIs "lousy filtering". In my last posts, I always said that it rarely happens, and that ATI is not generally inferior to NV, although I DID say this at first (which was wrong, my apologies). You're right about the conclusion, it was wrong to say that it is general inferior, but it's not better than NV.
And to be honest, you seem to be a little green when you say that all major sites would be on it already. Why do you believe that F1 2010 generally scored between 80 and 90 (on the given scale) despite being infested with bugs and using fake AI times (which is an absolute no-go for a racing game like this).
Yes, to most hardware magazines and sites, the banding is not worth mentioning, because it's not something that pops up instantly. But to me, that does not degrade its importance that much.
And funny thing is: No one here actually complains about the way AT tested ATIs AF. They didn't even show game content, all they showed in the article was that perfectly round AF-Flower. But that, of course, says a lot more about filtering quality!

@Aristotelian:
So you're a native German speaker? There IS potential for misinterpretation in that previous nick, and 99% of all members in 99% of all communities I am a member of are probably aware of it, and most communities are English communities.

 

[Aristotelian]

@Aristotelian:
So you're a native German speaker? There IS potential for misinterpretation in that previous nick, and 99% of all members in 99% of all communities I am a member of are probably aware of it, and most communities are English communities.

My language levels are 'native' in a few different languages, but I don't like to give too much information about myself out online. What do you have against the publishers?

 

[Blastman]

test

 

[KARpott]

My language levels are 'native' in a few different languages, but I don't like to give too much information about myself out online. What do you have against the publishers?

Sorry, but what publishers are you talking about? (maybe you just pm me if you're really interested, otherwise it's getting off-topic and the thread might be closed early)

 

[Idontcare]

There was nothing to misinterpret. So no.

@Aristotelian:
So you're a native German speaker? There IS potential for misinterpretation in that previous nick, and 99% of all members in 99% of all communities I am a member of are probably aware of it, and most communities are English communities.

Folks, you are free to have this seemingly private discussion in public if you like but typically this manner of discussion would be handled off-line between the two of you via private messaging (pm) or publicly in the Personal Forum Issues sub-forum.

 

[Blastman]

Games in which I observed the effect: TrackMania, Ut3, MirrorsEdge, Doom3, Prey, MaxPayne2, GTA4, Crysis, EliteForce2, HalfLife2, and I think there are many more, so I count at least 10 games, 8 more or less different engines and 3 APIs. I can show (and have shown) you screenshots if you like, many can be found in that 3DCenter-Thread I have linked to. Do you think they all have the same problem with the driver, on totally different engines and APIs, and why is it that it should only happen with games I have bought? Did you even follow the whole discussion here?

I've actually read through both the Beyond3D discussion and the one posted by BFG10K on his site, but came away unimpressed by what I saw. I saw no real indication that this was a "major" IQ problem affecting games and wrote the problem off as a bug in the AF tester being used.

Why should it load a different texture with 16xAF when it loads the right texture with 2xAF and only "partly" loads the right texture with 4xAF?

A fair point, so I went back and rechecked some of the screenshots posted. And I did find the issue you are relating to in some the screenshots posted -- although, even when you know what to look for, it's almost impossible to spot (unless you enhance the images). So, I'll post what I found and some brief comments.

First, you can use the filter effects in MS photo editor (sharpen and edge-thin) to contrast the images and check for bilinear mipmap transitions. I dug up some old images from SOF2 (Soldier of Fortune 2) to illustrate this (they were done on my old 9600XT).

SOF2 bilinear
SOF2 bilinear sharpen filter
SOF2 trilinear
SOF2 trilinear sharpen filter

As you can see, with the applied filter, the bilinear mipmap transitions really show. The trilinear filtered shot shows no harsh transitions. The edge filter (thin) works well too to highlight mipmap transitions.

On to some of the game pictures posted.

Crysis : original picture.
crysis

Filtered pictures.
crysis - sharpen 6 twice
crysis - edge (thin)

In both those filtered shots one can see a small mipmap band on the left side of the road (red arrows). But there is no banding anywhere else and the band that is evident only covers a small portion of the screen, limited to the tar/gravel textures on the side of the road.

GTA IV : original picture.
GTA IV

Filtered picture.
GTA IV - edge + sharpen 6

There is a slight banding on the left side of the sidewalk grate. I note that it doesn't extend across the screen to other textures. This one was hard to spot with just the edge filter, so I used a sharpen filter too. (note updated and removed street (left) red arrows as this appears more to be a texture change and not a mipmap line).


Wolfenstein : original picture. (Posted on page 2 of the German thread).
Wolfenstein

Filtered picture.
Wolfenstein - edge + sharpen 8

Nothing to note in this picture at all -- no banding evident, just shadows in the circled portion of the original.

UT3 : original picture.
UT3

Filtered picture.
UT3 edge + sharpen 6

Small banding area is limited to grate.

Judging by what I've seen so far, there is a small amount of banding on some textures in some games, but it appears quite limited -- and it only seems to appear on the first mipmap transition. In the case of Wolfenstein, there was nothing I could detect and the poster appears to be circling shadow transitions and thinking they are mipmap transitions. On this point, these things can get hyped up on the web up and we could get users thinking they're seeing bilinear mipmap transitions in all kinds of places where there are really not, when they don't know what to look for.

It may be that some recent optimizations have been pushed a little far, or perhaps ATI's AI was broken with a recent update and it is having problems on a few types of textures. It certainly doesn't appear to be a major IQ problem, but rather a minor IQ irritant, although granted my assessment is somewhat limited. I might consider doing some more extensive investigation of my own, but I don't currently have a late model ATI card. With the intense competition between ATI and NV one could certainly see each of these companies pushing the limit on optimizations right to the edge of acceptable. But it's hard to believe ATI would get much of a performance improvement from underfiltering a small portion of the textures on a grate or the side of the road, so I'm inclined to think at this point that something may be broken in the AI routine. But even given that, I'm simply not seeing large obtrusive mipmap bands in those games (with the exception of TrackMania), and nothing in Wolfenstein.

It's hard to say what's really going on here as these things only seem to be on a small portion of some of the screen shots, and they appear to be mipmap transitions. I'm not sure if higher levels of AF change the loading calls for the textures (I'm unsure of the technical aspects here), but that different grate texture in UT3 is certainly puzzling.

On the other side of the fence you have NV. When I was posting a lot in the food fight …er … graphics forum :biggrin:, a few years back, NV had worse issues than ATI with texture aliasing (which can't be shown in still shots). Given the comments I linked to earlier on nvidia.com, it appears that NV still has a worse problem than ATI in this area. So, at this stage, unless ATI fixes this, it might be a case of pick your poison -- worse texture aliasing in some games on one side, or a few small mipmap transitions in a few games on the other.

 

[BenSkywalker]

The artifacts offer a visual appearance along the lines of bilinear- it should *not* be confused with actual bilinear as it is happening on the *same* mip level- mainly mip 0. If it were an issue of actual bilinear then it would only happen on mip transitions- that is not what we are seeing.The only thing that could cause this is underfiltering outside of what the hardware should be doing(what the software asks for). Seeing evidence of something resembling bilinear in the middle of a given mip level shouldn't ever happen.

 

[KARpott]

@BlastMaster
It pleases me to see that, now, you are as far as I am. Yeah, I did say that ATI has inferior filters, which was wrong, and I apologized. And I DID say that this IS NOT a MAJOR issue. And many guys here said that NV is not perfect as well, which is true imo.
Usually, game developers try to avoid textures with high repetition frequencies and lots of colour contrast, because those textures are known to be problematic for texture filters. As you have correctly recognized, the banding strongly depends on the texture content, a lot more than the problems we saw with g70. It's simply NOT EXISTANT when looking at the so-called average ingame texture. To me, this indicates either a very clever AF-optimization which is not visible in 95% of all cases or some kind of hardware limitation which requires the card to somehow degrade the filtering when struggling with these textures.
The first case would, as such, not be bad at all. As long as the optimization does not degrade image quality, it's welcome. In case it does degrade quality, I would at least want to have some kind of checkbox that disables this optimization, as it has happened with the A.I. thingy.
The latter wouldn't be that great after all, as we don't know if solving the problem would have a great impact on performance.
There's one point why I can't believe in a bug/broken A.I. routine. The problem theoretically has existed since the HD4K release, as the aftester pictures indicate (although no banding is visible in games with HD4K). We all know about ATI drivers, and their, let's say, lack of perfectness. But if there's just a small routine that does not work at the moment, why wouldn't they have fixed it by now? I mean, if this would just be "one afternoon with a little debugging and code fixing", the costs would be minimal, considering that they could please all the IQ freaks with perfect texture filters in every situation.
To me, this sounds more convincing: ATI is aware of the issue and knows what's wrong. But as most of you, they don't think this "minor" problem would be worth fixing it, because that would cost money and probably degrade performance to an extend where they believe that it would be smarter to please the people with longer fps bars, because that would give them good press as sites like AnandTech obviously don't look too much at the details of the real world (I mean games ). And quite frankly, I believe the fact that they didn't respond to BFGs findings is the perfect confirmation, although I clearly have to state that this is JUST MY OPINION. And that leads me to Carsten again, whom I quoted I several posts above.

@BenSkywalker: Are you sure that it happens on the same mip-level? The lines are displaced (relatively to simple bilinear filtering) a bit because with 16xAF the lod is adjusted as well. I think that the UT shot actually speaks against this because the transition is so abrupt. And comparing bilinear with 16xAF in TrackMania, I'm actually quite sure that everything "behind" the transition belongs to the next miplevel, that's why I have actually included all the bilinear stuff. And isn't that what such optimizations usually do? Trying to find a middle-thing between bilinear and trilinear filtering? And why else should those line always coincide with the mip-transitions (apart from the different LOD)? Do you think that all the named games here ask the hardware to do some underfiltering, because it would be cool to have bad IQ on HD5K? Maybe some "the way it's meant to be played"-games would do it, but certainly not HL2, which is known for its slight affinity to ATI products, at least in the past. Please correct me if I am misunderstanding your argument, but that's how I have understood this "what the software asks for".

 

[BFG10K]

Well, sgssaa for eliteforce 1,2 and for rtcw does not work with Cat 10.9 here, but it works with the OpenGL HD5Krelease driver. All guys from 3DCenter have the same problem, the tool has been extended specifically for re-enabling sgssaa in older OpenGL games (that's why the release driver files are included), so I have no idea why it works in jedi academy etc.

That’s very interesting, and it must be a recent development because I remember SSAA working with older drivers in all of the Quake 3 games I tried, which included SoF 2 and Elite Force 2.

Are you sure that it happens on the same mip-level?

Yes, the tester app shows it happens on the base texture (i.e. before the mip-maps). The artifacts look like bilinear AF, but something else is actually the cause.

 

[BFG10K]

I think you’re going to have to make a much better case than an isolated instance (or 2) of lousy filtering on a couple of grates and Trackmania (where something is evidently "broken" and thus is not representative of ATI's filtering quality) to make the case that ATI’s filtering is generally inferior to NV’s.

These are not isolated cases; they happen in a large range of games, even in games that explicitly set trilinear like Doom 3.

Furthermore, ATi's 4xxx series does not show banding in the same situations.

If these bilinear transitions were a common problem and not something broken in a few rare instances, you could be sure the gaming sites would be all over this like a dung beetle on a fresh pile of poop as a major IQ problem (as I’m sure Nvidia would let them know about it).

Most gaming sites never picked up that ATi's 4xxx series had noticeably more aliased AF compared to nVidia. Or that nVidia's G80 launch drivers sucked utter balls. Or that the G7x's default AF quality was unusably horrific. Or that the GTX470 isn't really much faster than the GTX285 overall.

That’s why we have to sometimes go beyond mainstream sites to find answers, because it's end-users that regularly game on these parts.

It may be that some recent optimizations have been pushed a little far, or perhaps ATI's AI was broken with a recent update and it is having problems on a few types of textures.

These problems have been around since the September 2009 launch drivers (my article was written with CD drivers) and persist across different APIs, games and OSes (confirmed to happen with XP, Vista and 7).

That's more than a year; if it was a simple driver problem it would have been fixed by now, but all attempts to get a response out of ATi have failed.

This is a widespread problem that can’t be explained away on an individual basis, though obviously many won’t care and/or notice.

 

[Dribble]

I'm sure it's fixable. If for example it was to appear on the front page of [H] in a big article and Kyle bashed ati for it I bet it would be fixed (see xfire articles - [H]'s bashed ati there and they magically started working much better again in the next release). If someone big and important doesn't make a fuss it seems to take ati a long time to fix issues like these - I mean it took them the best part of a year to fix 120hz monitors for example.

 

[Madcatatlas]

Honestly i agree. There should be more critique from sites like Anandtech, Hardocp etc

Hardocps XfirevsSli reviews sure helped alot of people get a better gaming experiance in those 3 games.

It shouldnt be needed. AMD should have been doing that by its own accord. What exactly are those engineers/programmers doing after having released an entire series of cards. I mean yes, they probably are continously working on driver updates. But why didnt they see the huge potential for improvement until PRODDED by Hardocp and Kyle..?


edit: and ofc i forgot the topic comment: Is this the case here aswell? Is it hardware or is it software?

 

[MrK6]

Honestly i agree. There should be more critique from sites like Anandtech, Hardocp etc

Hardocps XfirevsSli reviews sure helped alot of people get a better gaming experiance in those 3 games.

It shouldnt be needed. AMD should have been doing that by its own accord. What exactly are those engineers/programmers doing after having released an entire series of cards. I mean yes, they probably are continously working on driver updates. But why didnt they see the huge potential for improvement until PRODDED by Hardocp and Kyle..?

It's because they didn't know about it until Kyle brought it to their attention. If you read the article (or maybe it's in the discussion thread linked to it), the drivers had passed all of their internal testing with no issue. For some reason, certain hardware combinations wreak havoc with their drivers. This also might be why some things get "broken" again presumably after they were fixed in earlier driver revisions. I'm no driver or software guru by any means, but it would seem to me this would be their biggest area of focus - not to break anything after it's been "fixed." Releasing a driver that doesn't do much isn't great, but it isn't bad; releasing a driver that breaks things that were working previously will piss people off, as we've seen.

edit: and ofc i forgot the topic comment: Is this the case here aswell? Is it hardware or is it software?

Both, I've heard. Some say it's a cache limitation in the architecture. It will be very interesting to see if the 6xxx series brings about any changes.

 

[KARpott]

That’s very interesting, and it must be a recent development because I remember SSAA working with older drivers in all of the Quake 3 games I tried, which included SoF 2 and Elite Force 2.

Indeed. With the latest driver, elite force 1/2 and I think many other OpenGL games suffer from lots of stuttering and a lack of good ssaa. I dunno if there are newer drivers than the old HD5krelease driver that also work, but in case you wanne enjoy these old pearls again (EF 2 was a short but great game imo, a lot of StarTrek atmosphere, a huge amount of litte gimmicks including a side-scroller mini game, just the kind of shooter I miss today), you guys might wanna give it a try with that tool I mentioned, in case you own a hd5k series card. A lot more comfortable than going completely back to the old driver.
Oh, and I think you're right about this mip-level thingy. Just took a look at the AFtester shots again.

 

[distinctively]

r.

EDIT2: Btw, just coming back from a flight through UT3. I bet you have not touched the map "Defiance", did you? I can perfectly reconstruct the UT shot in my first post. But I agree with you that 99% of all textures in UT do not have visible transitions, although I actually found another texture with the same problem (different to the one from the shot above).
Hint: Look for heavily shaded textures with VERY VERY high repetition frequencies. It's very likely that you can witness the "banding" on VERY VERY VERY fine grates, with a lot of contrast. As I said before, the visibility not only depends on the texture itself, it also depends on how it is shaded or lit.

Took a while. I'm very busy at work. I looked at the UT3 defiance as you suggested in the same area. There is a big skylight window above. The scene changes with every single degree of movement. I can make the transitions visible or invisible. It just depends on which way the light is hitting the grate below. Gimme a bit of time and I'll load a few shots. I'm just swamped with work this month.

 

[KARpott]

Gathered some more footage which might bring a little more light into that case.

I used the D3DAFTester for the following pictures and vids:
http://www.3dcenter.org/download/d3d-af-tester
You might wanna visit this site again for explanations:
http://en.wikipedia.org/wiki/Anisotropic_filtering

D3DFilterTester-Comparison:

Settings:
ATI: HD5850/Cat 10.9 - 800x600 - no AA - AF via application - A.I. off - MipMap HQ
NV: GTX260/FW258.96 - 800x600 - no AA - AF via application - driver HQ (thanks to Airbag from 3DCenter)
NV2: GTX470/FW258.96 - 800x600 - no AA - AF via application - driver HQ, clamp off (thanks to kruemelmonster from 3DCenter)
Application settings: LODBias 0, mipmap: OFF - Use Stage: 0 - Use Checkerboard Texture/scale 0 - Objects: Plane/Distance: 2.2/Angle: 57 - scene/texture rotation: 0.0

At first, a few ATI screens


The colors represent the number of texel-samples taken for one pixel. Green is the maximum, red is the minimum.
As you have stated, the problem is not dependent of mipmaps, as mipmapping has been disabled for these pictures. But the transitions in the real scene coincide with the transitions in the analyzed scene.

Now the comparison shots:
---------------Trilinear---------------------------------16xAF---------------


Until now, it's not entirely sure if the scales for ATI and NV are actually the same, but one thing is clear: The NV transitions are much smoother whereas ATI filters seem to be working more angle-independent. Btw. , the analyzed pictures do not depend on the actual texture content, that just means that, for obvious reasons, the texture filters do not react dynamically to the texture content.

D3DFilterTester Video-Comparison

Additionally, there are some videos I made. The LODBias was shifted from 0 to 8.
ATI: http://www.multiupload.com/I6TCASAGZ4
NV: http://www.multiupload.com/LKG279KT6K
Settings:
ATI: as above, with mipmaps: linear, normal non-colored representation!
NV: GTX460/FW 259.57 - driver HQ - same as with ATI (thanks to Gorkon from 3DCenter)

Now with texture-scale 8:
ATI: http://www.multiupload.com/XBD7GZP1N0
NV: same file as before

In case both scales for ATI and NV are the same (I will find that out soon), I would draw the following conclusion:
To me, those pictures actually show that ATI does not generally underfilter. Actually, it even seems to take more texel-samples at many points (given that both scales for NV and ATI are the same, see above), but the transitions are very sharp. I think that this causes the banding. Additionally, there could also be a problem with the Texel-Sample pattern. I mean, look at the very sharp transitions in the ATI video. I doubt that, on the lowest mipmap-level with the least texels, such transitions would just originate from abrupt changes in the number of samples taken for the given pixel.

 

[TLDR]

Might be an interesting clue on the source of the problem. Take a look at this post:

http://forum.beyond3d.com/showpost.php?p=1479306&postcount=75

 

[Xarick]

So basically both karpott and the link seem to point to ATIs algorithm as the problem.

 

[KARpott]

The forum entry at beyond3d actually points directly at what I have noticed as well.

 

[KARpott]

If not faked, this seems to be the ultimate response from AMD to our issue:
http://gathering.tweakers.net/forum/list_message/34877928#34877928

 

[HurleyBird]

Thank God. Would be nice to finally have "perfect" AF, since I think this is one area that was a significant downgrade coming from my G80 to my HD5870. Would be nice to have a high resolution image of this screen so we can interpret it better.

 

[SirPauly]

It's great to see.