ATi 5000 Series Image Quality Analysis

The GTX 780 arrives

– 25 Games benchmarked!

Posted on November 1, 2009 by BFG10K Posted in Articles, Technology 45 Comments

Anisotropic Filtering Theory

I’ll start with anisotropic filtering. In past articles I demonstrated that ATi’s AF was inferior to nVidia’s because it suffered from more aliasing and shimmering. The 5xxx series offers improved AF over ATi’s previous generations, so it’ll be interesting to see how it looks in practice.

The main improvement ATi has brought to the table is angle invariance. This essentially means that all surfaces are filtered with equal strength, regardless of their angle relative to the camera. This improvement could be partly possible due to the removal of the fixed function interpolators that bottlenecked performance and lacked precision, and by moving some of the calculations onto the shaders. Aside from this enhancement, I’m also keen to see if any improvement has been made to surface filtering strength, which is the main area where ATi was visibly weaker than nVidia in the past.

First, here’s AF applied to a theoretical rendering of a tightly textured black and white pattern. These pictures were taken with 16xAF and 8xMSAA (click them to see larger versions).

4850

285 AF 150x150 ATi 5000 Series Image Quality Analysis

285

5770

We’ll start by paying attention to the colored areas that begin with the red section and move into your screen. This area represents mip-mapping, which is smaller textures derived from the base texture.

With the 4850 we see the red and green areas are square, while the other colors are clover shaped. This is actually the worst implementation of the three cards, but it was better than current implementations at the time it debuted on ATi’s X1xxx series.

With the GTX285, we immediately spot the huge improvement over the 4850. The red section is the worst but given it’s octagonal in shape, it’s still much better than the 4850’s box. While the rest of the colors still have a slight flower shape, they’re far closer to being circles than the same 4850 colors are.

With the 5770, we see all of the rings are circular without any hint of flower, clover or square shaping. This includes even the tough red section that the other two cards have so much trouble with. Thus the 5xxx is the first board in consumer space to offer truly angle invariant anisotropic filtering at a 16x level.

Next we’ll check filtering strength, and we can do this by ignoring the colored areas and instead focusing on the black/white/gray areas in front of them. The star shaped patterns are texture aliasing caused by undersampling. This texture aliasing is responsible for flickering and shimmering on textured surfaces during gaming.

With the 4850, we can see more stars appear as we move towards the center, and as we touch the red section, we see multiple adjacent stars. I’ve marked two of them with white arrows.

The GTX285 also has the same issue, but as we move further back and touch the red section, the 285’s stars are far less pronounced than the 4850’s. If this were a moving scene, these sections would shimmer less on the GTX285 because the texture filtering is stronger than the 4850.

Moving on to the 5770, we see something bizarre happen. It introduces sharp gray transitions of varying texture detail, eventually culminating into what can be called a “dead-zone” – the solid gray ring touching the outer red section with no discernable texture pattern applied to it. Furthermore, the largest gray ring introduces new stars that the other two cards don’t have (two such instances are marked with white arrows).

While there are no stars in the “dead-zone” – and hence that area should shimmer less compared to the other cards – there’s also no obvious texture pattern applied to it either, so this solution isn’t correct. In actual gaming, I’d expect this area to be blurrier than the other cards in comparison.

Based on these results, 5770’s angle invariance is outstanding, but its surface filtering is unacceptably poor and is worse than the other two cards. I can only hope it’s a driver bug that can be fixed.

Incidentally, here’s the same scene with 8xSS on the 5770:

5770 AF 8xSS 150x150 ATi 5000 Series Image Quality Analysis

5770 (8xSS)

With 8xSS, the surfaces in the gray rings look absolutely superb, and all star patterns are almost invisible. Sadly the sharp transitions are still present however.

Pages: 1 2 3 4 5 6 7 8 9 10 11

« AMD Athlon II X4 630 Extreme Overclocking with LN2 by Deanzo

SparkLAN CAS-673W Wireless Pan and Tilt MPEG4/MJPEG IP Camera »

Bo_Fox

Wow, this is a godly article that I have been looking for, for a long time. PCPer had one that showed the differences of adjusting LOD with SSAA in Crysis, with -2.0 LOD lessening the blurriness of 8x SSAA, but I’ve never seen such an in-depth analysis of the hidden flaws of ATI’s new perfectly circular AF method!!!

One other thing to note about Nvidia’s AF method is that in the Nvidia Control Panel, one could select “High Quality” and turn off AF and other optimizations. That way, we are almost guaranteed to see truly trilinear AF. There is no way in ATI’s CCC to disable those optimizations (unless one completely turns off Catalyst A.I., which could also remove some of the valuable optimizations for a certain game). Well, I guess the hidden transitions of varying filtering detail is still not fixed by disabling Catalyst A.I., but at least it should remove some of the “Brilinear” AF culprits.

Thanks a bunch, BFG10K, for bringing this to light. I hope ATI takes note of this and realizes that it needs to be fixed, especially after the public now knows about it! I remember you from the Nvidia forums. Keep it up, BFG10K.. great work!!
Bo_Fox

Oops.. I mean PCGamesHardware (not PCPer). Also, I mean AF optimizations (not turning off AF altogether). Perhaps editing it for me would be appreciated? Thanks in advance.
BFG10K

Bo_Fox, I always operate nVidia hardware with high quality filtering, all optimizations off, and LOD clamp enabled. Any results of mine always have the highest image quality possible on nVidia (and ATi) cards.

P.S. you should be able to edit your own posts. To the right of the date/time stamp of each post there’s a small “edit” button.
Bo_Fox

No, no, I was not directing it at you, BFG10K!! It was only to complement your article with additional insight. Of course, you knew better.. I would’ve had to be dumb to think that you were dumb.

This is a great article, BFG10K! I do not know if it was you who originally discovered this “brilinear” AF (dare I say, cheat?) problem with ATI’s new circular AF method. I think that a lot of the 16x AF blurriness has to do with the fact that none of the texture patterns are visibly shown in the red area (the outer transparent areas of red). Does it make sense to you?

Once again, great article.
Bo_Fox

I can’t find the edit button next to the date/timestamp. It’s nowhere to be found, even after highlighting everything on the page. Perhaps only the moderators have this ability?

I really hope ATI fixes this problem! Thanks once again for bringing it to attention.
Bo_Fox

Hi BFG10K, I’m wondering about the AF image quality with newer Catalyst drivers.

I’m also wondering if this “Brilinear” IQ only applies to 57xx cards, not 58xx cards? When I first saw reviews of the 5870 GPU’s, it showed a nice, clear AF with smooth progressions (the texture checkerboard pattern showed behind the outer edges of the red transparent area). Here’s a link to the picture: http://images.bit-tech.net/content_images/2009/09/ati-radeon-hd-5870-architecture-analysis/aniso-filtering.jpg

Or it could be possible that ATI has dumbed down the IQ about one month later after the 5870 was released, after all the “big” reviews were done (for “so-called” driver speed enhancements)?
Bo_Fox

Also, what program did you use for testing it with this fine moire checkerboard pattern? I have D3DAFTester, but have never seen such a detailed background.
BFG10K

Bo_Fox, the screenshots you linked to were taken with the default texture pattern. This pattern is not tight enough to expose limitations with AF. I use AF tester too, but I tighten the pattern significantly.

I have no reason to believe the 5800 is any different to the 5700 in this matter. Also my tests were done with the CD drivers, which means the problem has been right there from the start, but other reviewers didn’t seem to notice it.

No, I haven’t retested 9.12, but in a few months I hope to retest many games’ performance, and also check on the AF issue.
Bo_Fox

Thanks, BFG10K! So far, you’re still the only one to have ever tested for this particular AF problem with some screen-shot evidence. I’ve tried spreading word about it on the forums, and this is the reason why I have not yet bought HD5k series for myself. (I bought a 5750 for my younger brother, but it was not compatible with his Nforce4 chipset so I returned it for a 4890). On the forums, people with HD5k cards still do not want to bother testing their cards for this AF problem with the newer drivers to confirm with me on whether this issue still exists or not.

I hope you can check upon this issue once again real soon, especially if you have a 5870 card. If this brilinear problem is confirmed to exist with 5870 cards, it will stir up quite a commotion among the owners of 5870 cards who have paid dearly for such. The more pictures you can provide using as many games as possible, the better. Also make sure to provide a screenshot of the exact settings used in Catalyst. I’d love to see your reputation take off quite a bit, becoming known as the man to point out this problem, if it does exist (which is so contradictory with ATI’s marketing of a perfectly circular AF method).
KrossX

@Bo_Fox: That screenshot comparison doesn’t show the issue since the checkers are too big.

To get the same tiny checkboard, just move the slider under Textures. Using -2 seems to be just like on this article.

@BFG10K: Great article. With that AF comparison, you gave me the push I needed to actually at least post about it. (and use the feedback form to AMD)
titaniumx3

Just tried this myself on a HD5870 with Catalyst 9.12 drivers and I too get the so called ‘deadzones’ just like it is in the article.

I’m a little suprised to say the least and although I appreciate that the effect is far less obvious in games, it’s still concerning, especially for a high end card.

I’m wondering if this is a fixed hardware issue or if it is possible to change it via the drivers. I very much it is the latter and that enough people raise this issue for it to be taken up by the Catalyst driver team.
KrossX

@titaniumx3: If the game has the option to force trilinear filtering, that fixes it for the game. ATI Tray Tools also offer the option to Force Trilinear globally, so I don’t think it’s too much of a driver issue… rather than some setting.

Please do use the Catalyst Feedback form to report the issue. And if you have an account on the AMD forum, check the following thread:

http://forums.amd.com/game/messageview.cfm?catid=260&threadid=126249
BFG10K

KrossX, nice to see you spreading the word.

After I wrote the article, I actually emailed Terry Makedon (Catalyst Maker) directly with my findings.

Also trilinear filtering doesn’t necessarily fix the problem; the Doom 3 screenshots were taken with the game running under trilinear, yet you can clearly see the transitions in the screenshots.
KrossX

That’s not nice. So far, enabling trilinear has been working fine in the few games I’ve tried it.

At least they should know now. So that’s good. ^_^
Tommy Haegin

Hello BFG10K,
this was a most informative article. I’ve only discovered it today via a link I’ve seen posted in a forum. Do you still plan to revisit the AF issues on the 5xxx cards sometimes with newest Catalyst drivers? I would much appreciate.
Thank you very much again for this revealing article!
Tommy
titaniumx3

Just tried the Catalyst 10.2 drivers and still getting the filtering deadzones

Also it may just be in my head, but I’m beginning to notice these deadzones in certain games and it’s pretty distracting, like sudden jumps in blurriness.

I hope more people speak up about this, because personally I’m a little annoyed that my previous generation HD 4890 never exhibited these problems, yet the HD 5870 does.
BFG10K

Tommy Haegin:

Yes I’ll revisit image quality and performance on the 5770.

titaniumx3:

I don’t think it’s in your head at all. Once you know what to look for, you’ll begin to notice such problems. It’s a lot like aliasing; jagged edges are as plain as day if you try going without AA.
http://scalibq.spaces.live.com Scali

Ofcourse the gray area is correct for the Radeon 5770.
It’s a checkerboard texture, so the smallest mipmap will be a 1×1 gray pixel.
You could only get any kind of pattern if it was undersampling (and thus sampling from a mipmap of a level too high). Else, you get gray, as it can’t go beyond the smallest mipmap.
BFG10K

Scali, you must be joking.

The gray area is *absolutely wrong* because the 5770 is undersampling, so it’s unable to texture the surface with the correct detail. The sharp transitions are also wrong, as is the increased texture aliasing in the outer ring.

The extra blurriness and transitions from the 5770 are visible in screenshots, and are even more visible during in-game movement.
Scali

How do you reckon it’s undersampling?
As I said, if it were undersampling, it would NOT arrive at gray. I’ve given arguments, and no I wasn’t joking. You are not producing any counter-arguments.

The others are aliasing and painting moire patterns, which makes it APPEAR that there is more detail, but in fact there is not, it’s just aliasing. This also hides the transitions (which are a result of the linear filtering between mipmaps, and not directly related to anisotropic filtering).
BFG10K

The transitions are from bilinear-esque optimizations at play, and are quite visible during in-game movement. If you think that’s normal then you don’t understand filtering. Also the 5000 has more texture aliasing in the nearest section, proving yet again that it’s undersampling.

The fact that the 5000 starts painting solid grey even before it hits the first mip-map also proves it’s undersampling. Yes, given a far enough distance things should look greyish, but not solid like that. The solid section is devoid of any detail whatsoever, and is far too close to the camera.

Any elementary paper on filtering demonstrates what a properly filtered tiled floor looks like. I’ll leave that as an exercise for you to do some research on the matter.
Scali

Why would they be bilinear-esque optimizations? You realize that filtering between mipmaps is only linear, don’t you? Anisotropic filtering only applies to minification and magnification filters. So you always sample two mipmaps at best. At the least that explains SOME of the transitions, as clearly you will see a transition when you sample only from the gray mipmaps, and also when you go from two black/white ones to one gray and one black/white one.
The transitions are more obvious on the 5000 series since they are perfectly angle-independent, which means the transition occurs at the same distance in the cylinder for every pixel at every angle, leading to the circular patterns. If you are more angle-dependent, you will not switch mipmaps at the same place everywhere, and ironically enough the transition becomes less obvious.

And how do you reckon the solid gray is not valid?
The checkerboard is set to a pattern of 2 pixel-wide squares… Now I don’t know if that means that they actually use a 4×4 texture, or a larger texture with repeating 4×4 blocks… but for the rest of the story, that is not relevant.
Namely, you can create exactly ONE smaller mipmap that is still black/white, that would be a texture with 1-pixel-wide squares. Obviously any smaller mipmap will result in a completely gray surface.

Looking at the progression of the checkerboard from near to far, I think it’s perfectly obvious that the gray starts to occur exactly where you expect it: at the point where the 2×2 pattern of pixels needs to be fit into a single pixel. The gray mipmap starts to ‘bleed’ into the image.

QED
Scali

I’d like to add that this is specific to the way filtering is implemented in D3D/OpenGL. They put anisotropic filtering before mipmap filtering, and as such it cannot be perfect by definition.
So ‘elementary papers on filtering’ are nice theory, but they may not necesarily be relevant to a D3D/OpenGL application. There are tons of filters (even anisotropic ones) that have less aliasing than a reference D3D/OpenGL implementation.
BFG10K

Why would they be bilinear-esque optimizations?

Because that’s exactly what you get when you you’re performing bilinear AF, namely the sharp transitions. Anyone with a basic understanding of AF knows that. That and forcing trilinear AF makes the problem go away in some games, as is testified by some people in this thread that have actually tried it.

Namely, you can create exactly ONE smaller mipmap that is still black/white, that would be a texture with 1-pixel-wide squares.

The section that has the dark grey isn’t even a mip-mapping section and if it was, it wouldn’t be at such a small level. Furthermore, the first band has more texture aliasing on the 5770 compared to the other parts, but you don’t seem to understand this.

If you are more angle-dependent, you will not switch mipmaps at the same place everywhere, and ironically enough the transition becomes less obvious.

If the other cards had the same problem then we’d still see square or octagonal transitions, but we aren’t.

So ‘elementary papers on filtering’ are nice theory, but they may not necesarily be relevant to a D3D/OpenGL application. There are tons of filters (even anisotropic ones) that have less aliasing than a reference D3D/OpenGL implementation.

The 5000 is further away from perfect AF compared to nVidia’s current parts. The way you describe the 5770’s AF deficiencies as “correct” and “better” is completely wrong.
http://scalibq.spaces.live.com Scali

What is “bilinear AF”?
Bilinear anisotropic filtering? It’s either bilinear or anisotropic filtering, not both.

And “trilinear AF”?
Trilinear filtering means bilinear minification and maginfication, and linear mipmap filtering. Again, it’s either trilinear or anisotropic, not both.

I think you have to be more specific. D3D and OpenGL support three types of filters:
Minification, magnification and mip.
You should be more specific and mention all three filters if you want to describe what filters you’re referring to in an unambiguous fashion.

Right now you’re just contradicting yourself and confusing the issue.
BenSkywalker

Where to start, perhaps the most recent comments. You state it is either trilinear or anisotropic, not both. Microsoft disagrees with you-

http://msdn.microsoft.com/en-us/library/bb172264(VS.85).aspx

You set your map parameter to either point or linear, if it is set to point you end up with bilinear, linear gives you trilinear.

The interesting thing to note about both the clear sampling reduction and the graying out of the area is that both occur prior to the first mip transition in the D3D AF tester. The colored bands visible are the mip transitions- all of the visual oddities happening in that test are happening on mip 0. That seems to be where your understanding of this test is failing you the most.

I find it highly unlikely you are going to listen to reality here though, perhaps we should take this discussion to B3D and they can hammer out your inability to comprehend this simple test.
BFG10K

It’s either bilinear or anisotropic filtering, not both.

This is absolutely untrue. When performing AF you can have bilinear or trilinear sampling.

If you employ bilinear you’ll get sharp transitions in the mip-maps, similar to what is happening with the base texture in the test. Anyone experienced with older parts (e.g. Radeon 8500) knows what bilinear AF looks like.

But I specifically said “bilinear-esque” because the 5770′s problems are happening at the base texture, so while it’s quite similar to bilinear AF, it’s not exactly the same thing.

You don’t understand the test and you don’t understand AF, yet you continue to argue the topic. This is exactly what you were doing when you were arguing MSAA about a year ago.
Scali

“This is absolutely untrue. When performing AF you can have bilinear or trilinear sampling.”

No it’s not. Bilinear and trilinear sampling are isotropic by default.
Bilinear and trilinear are common names for certain sets of filters, and are not actual filtering options in the D3D API.
Bilinear is when you sample the nearest mipmap linearly in two dimensions(being u and v). Linear filtering is isotropic by default.
So the bilinear case uses a linear filter for minification and magnification, and nearest neighbour for mipmaps.

Trilinear is when you take two bilinear samples from the nearest two mipmaps, and then linearly filter those two samples, hence you do linear sampling in three dimensions, namely u, v and miplevel.
In other words, minification and magnification are linear filtering, and mipmaps are also filtered linearly.

Anisotropic filtering can only apply to minification and magnification, and can be used instead of the linear filtering in u/v directions.

I think what you’re trying to argue (but without understanding the actual API and terminology) is that AMD applies nearest sampling for the mipmapping only, rather than linear filtering, under some conditions. This however would show up as clear banding in the coloured mipmaps, but we don’t see that happening (not even when you modify the LOD bias so that the coloured mipmaps cover all of the screen).
Scali

I’d like to add that linear mipmap filtering is commonly assumed when AF is enabled, as AF is considered to be one step above trilinear filtering.
But strictly from a D3D/OpenGL point-of-view, this is not a requirement. As I said, you can set three types of filters indepedendently, minification, magnification and mip.
If you set min and mag to AF, and mip to nearest rather than linear, is it any less AF? Mipmap filtering does not depend on isotropy, as you filter only two samples, with no direct relation between the shape of the pixel on screen (which was dealt with in the min/mag stage). Therefore isotropy has no meaning for this particular filtering stage.
Scali

“http://msdn.microsoft.com/en-us/library/bb172264(VS.85).aspx

You set your map parameter to either point or linear, if it is set to point you end up with bilinear, linear gives you trilinear.”

That page doesn’t mention ‘trilinear’ at all. So it doesn’t prove anything, except that you understand about as little of the D3D API as BFG10K does.
‘Bilinear’, ‘trilinear’ and ‘anisotropic’ are jargon invented because you don’t want to explain all three filter stages everytime.
I ask you this: how can you have trilinear (tri meaning 3) filtering if only one of your stages is linear filtering?

Really guys, you can drop the attitude, you’re out of your league here. Feel free to take it to B3D, and they’ll explain that I’m right.
BenSkywalker

“It’s either bilinear or anisotropic filtering, not both.”

You are in no uncertain terms wrong even using the very narrow case of explicit D3D calls. You can try and backpedal and say that it is assumed, but it is clearly spelled out in the most elementary of documentation. If your map is set for point, you are getting bilinear filtering. You can try to agrue semantics based on current D3D calls, that simply demonstrates how young you are as these filtering techniques predate these particular D3D calls by quite some time. Neither BFG nor myself are stating that ATi is using bilinear filtering, we are commenting on visible artifacts comparable to what bilinear filtering gives which brings us back around to the more important element in this discussion.

The ‘gray’ area is mip 0. How does this compute with your assertion that it is correct?
BFG10K

Bilinear and trilinear sampling are isotropic by default.

Yeah, and? That doesn’t stop them from being in effect when AF is being used. You still have to define how those mip-maps are sampled, and bilinear AF means sharp transitions because there’s no linear blending between them.

This is really quite basic: http://www.beyond3d.com/content/reviews/37/16

The screenshots at the bottom show trilinear, bilinear AF, and trilinear AF. If you can’t understand those pictures then I don’t think anyone can explain this to you.

You also don’t understand that the 5770 has more texture aliasing in the outer ring, because I suspect you don’t even know what texture aliasing is, or what it looks like.

Really guys, you can drop the attitude, you’re out of your league here. Feel free to take it to B3D, and they’ll explain that I’m right.

This isn’t your blog; here you don’t get to make nonsensical comments and not get a response.
Scali

“The ‘gray’ area is mip 0. How does this compute with your assertion that it is correct?”

It’s not mip 0. The texture is a lot larger than the pattern painted on it. The texture pattern is set to ’2′, which as far as I can tell means that it uses 2×2 squares of black and white pixels.
As I said already, you can have only ONE mipmap smaller than this largest texture, and still have a black/white pattern. Every smaller mipmap will simply be gray, and that could affect a lot more mipmaps than just mip 0, depending on how large the original texture is.
Scali

“You still have to define how those mip-maps are sampled, and bilinear AF means sharp transitions because there’s no linear blending between them.”

That’s my point, you are using confusing terminology.
Bilinear filtering could even mean that no mipmaps are used at all.
As I alraedy said, you need to specify what type of mipfilter is being used, if you want to be specific.

Aside from that, they *could* be using bilinear *samples* (not texture filter, samples), or trilinear *samples* as part of the anisotropic filter… But that is very implementation-specific… just like the number of mipmaps that their trilinear samples are allowed to use in total, for example.
Since nobody knows exactly how AMD implemented their filter, you shouldn’t be so adamant that you are right about this (especially since you can’t even get the terminology right).

All I’m saying is that if they were using nearest sampling for mipmaps in some cases, it would show up as solid bands of colour in the coloured mipmaps… which it doesn’t. Which you haven’t even commented on, so I assume that means you know I’m right on that matter.

But if you feel you need to insult me in order to try and make the point you don’t have, go ahead.
I don’t care whether or not I convince you. You’re just some internet punks with no actual knowledge or experience in the industry, with a big mouth.

I’m just posting here so the readers can get a more balanced view and make up their own minds.
Tamlin

I tested this myself with a ATI 5870. Unfortunately, I don’t have an adequate Nvidia card to test it against.

As BFG showed, there is a star bottom and top on the 5870 with 16AF which isn’t on the picture from GTX285.
The stars on left and right on the GTX285 isn’t as pronounced on the 5870.

With 16AF and mipmap on max, there is grey rings as shown in BFG’s picture. If I reduce the mipmap level and amount of AF, the rings goes away. If mipmap is reduced too low, the chessboard pattern gets too large, but rings doesn’t appear and there is pattern instead of grey rings.

I can change lod bias and the colored mipmaps changes size, but not quality.

They only way I can show chessboard patterns instead of grey rings is to reduce mipmap quality or the size of the chessboard texture (2 and lower gives grey rings and no pattern).

It seems my 5870 is incapable of showing pattern 2 and lower correctly. On the good side, texture seems sharper (GTX285 texture seems to turn more grey and blurry on the picture the closer to center it becomes)

Something seems wrong on texture 2 and lower with 5870, but above it seems great. I wonder if BFG could do a follow-up using more samples to see the AF of both cards in several scenarios?
BFG10K

That’s my point, you are using confusing terminology.
Bilinear filtering could even mean that no mipmaps are used at all.
As I alraedy said, you need to specify what type of mipfilter is being used, if you want to be specific.

I’ve been very clear right from the start when I referred to it as “bilinear esque” and how it’s similar to the current situation, but not exactly the same.

It was you who stated you can’t have bi/tri running at the same time as AF, which is outright wrong. I even linked to screenshots clearly demonstrating bilinear AF and trilinear AF in action.

Now you’re arguing some irrelevant tangent about bilinear filtering with no mip-maps just to hide how wrong you were.

Since nobody knows exactly how AMD implemented their filter, you shouldn’t be so adamant that you are right about this (especially since you can’t even get the terminology right).

Actually this goes back to what I said earlier about any elementary paper on AF demonstrating what a properly filtered floored looks like. It doesn’t have any such banding or solid gray areas that the 5000 does, and the result looks closest to what nVidia’s parts are producing in that test. In response to this, you started going down yet another irrelevant tangent about D3D/OGL.

Also anyone with any level of experience with these parts during gaming can see the visible texture transitions and reduced detail on the 5000. The documented proof is all right there in the screenshots I posted.

All I’m saying is that if they were using nearest sampling for mipmaps in some cases, it would show up as solid bands of colour in the coloured mipmaps… which it doesn’t. Which you haven’t even commented on, so I assume that means you know I’m right on that matter.

I haven’t commented on that because it’s not relevant. We’re not talking about the mip-maps, we’re talking about the base texture being sampled in such a way that it has banding artifacts similar to what a bilinear filter produces.

But if you feel you need to insult me in order to try and make the point you don’t have, go ahead.
I don’t care whether or not I convince you. You’re just some internet punks with no actual knowledge or experience in the industry, with a big mouth.

You were the one that was directly calling out both Ben and I in your blog, and linking directly to this article as if we were stupid or something. Don’t start fires if you can’t take the heat.

I may not be an expert but I understand enough of the fundamentals (and have enough practical experience in gaming) to know when someone is faking arguments by using big sentences full of irrelevant tangents to mask that fact.

I’m just posting here so the readers can get a more balanced view and make up their own minds.

LOL, sure you are.

Hint: correct AF isn’t up for debate because it’s mathematically provable. Again, any elementary paper on filtering will explain this.
BFG10K

Tamlin, thanks for joining us.

By reducing the AF on the Radeon, you’re reducing the amount of “damage” it does, so the artifacting is reduced. Of course by doing this the mip-maps shift closer to the camera since less area is getting AF.

While a negative LOD bias will move them back again, it’s a poor substitute for AF because it introduces additional texture aliasing. A negative LOD should never be used to correct AF because it means the AF isn’t working properly in the first place.

Compared to the other two parts, the 5770 is under-sampling the base texture in that test, and this starts to manifest itself when the texture pattern gets small enough.

BTW, as per your email question, I was using a LOD clamp with the GTX285, so it most certainly was NOT using a negative LOD bias.
Tamlin

Thanks BFG! I have followed your articles with great interest, but just recently tried out D3D AF tester myself.

Negative lod is something I tested out with to reduce the rings with colored mipmaps, so that the chessboard pattern would be more visible. Changing lod bias doesn’t seem to affect the chessboard pattern at all.

Try reducing the mipmap to performance in CCC (not high performance), start D3D AF tester and set the following:
texture 2, 2X AF, full colored mipmaps and reduce lod bias to -2 to see the chessboard pattern on a similar stage as the GTX285. The solid grey rings are gone and there is only a faint ring with grey tinting the pattern as with the GTX285 (gtx285 has a smaller ring though). If you use more AF samples, the grey rings are introduced again.

Close D3D AF tester and change the mipmap settings in CCC to high performance. Open D3D, select texture 2, 2X AF, tinted mipmaps and reduce lod down to -2 again. You’ll notice that the red colored mipmaps are tinting the whole chessboard pattern, but aliasing is greatly reduced. No grey rings. This time, the pattern are magnified.

It seems that the results are better if I undersample and lower quality of mipmaps, and that the size of the mipmaps are directly related to the grey rings.

Could you check this up against the GTX285 and see which behavior that card has with the chessboard pattern with altering the same settings?
BenSkywalker

http://technewshw.com/forum/showthread.php?t=2863

An interesting thread with some more tests showing the same issues.

“The texture pattern is set to ‘2?, which as far as I can tell means that it uses 2×2 squares of black and white pixels.”

This is just getting comical- you have never even run the application? I have it open right now set on 0- still no gray even on mip 9. A person with a bit of capacity may have went and downloaded the free app we are talking about and checked it out himself.

“You’re just some internet punks with no actual knowledge or experience in the industry, with a big mouth.”

Like the moron kid who said you couldn’t run bilinear and AF at the same time under D3D? Even after I linked the MSDN doc, you still try and spin it. I don’t think for an instant you have any experience at all- you can’t even recite back the most trivial D3D calls without getting yourself confused.
Tamlin

BFG: Found a new program that might shed some light upon this:
http://www.3dcenter.org/3dtools/filter-tester-en

Use TMU and ALU rendering in split screen, tunnel and you’ll see the behavior on the same sample with both methods.

Seems like the grey circles are coming on the TMU rendering, but not on the ALU rendering on my 5870.
BFG10K

That’s an interesting program there, Tamlin. AFAICT the ALU version is reference AF run in “software”, while the TMU version is the card’s native hardware implementation.

In that case it’s hardly surprising that there’s a difference between the two on your card.

On my GTX470 the two are almost identical, even with the ALU version set to 512xAF. Both have the same texture aliasing though, which proves even the strongest AF still needs super-sampling to clean it up.
Tamlin

Its not equal on the 5870. I made a post on B3D to see if they can explain what is wrong and if its broken somehow:
http://forum.beyond3d.com/showthread.php?t=57870
Tamlin

Check this out:
(in German, but pictures are pictures and they were made with an HD5700/5800)
It looks, like the new AI options (Performance -> Quality -> High Quality) influence the texture quality by adding more samples. But did AMD really improve the filtering quality? Or did they only decouple filtering chea… optimizations from game fixes, Crossfire…?
http://forum.beyond3d.com/showpost.php?p=1484429&postcount=85

Seems 10.10 beta with a reghack does some improvements.
BFG10K

There’s also a leaked slide that basically confirms my findings, and shows the 6000 series fixes it:

http://img217.imageshack.us/img217/7882/355595ds.png

Heck, they even use a similar tester app to mine to demonstrate the difference.

Oh, and check out my latest GeForce 400 image quality analysis article: http://alienbabeltech.com/main/?p=21497
army mos

Interesting blog, I’m going to spend more time reading about this topic