Explanation of ATIs AF advantage?

[BFG10K]

It is displaying what the anisotropic filtering implementations are doing.

Yes but it's displaying them in a way that no game does. I've never seen any game that has a single, flat, high-contrast colour wheel in the middle of the screen. Have you?

The excessive bowing and improper mip transitions are simply proof of what is evident in game.

No, it's evident if every game looked exactly like the colour wheel did. The colour wheel is designed to bring out the absolute worst case scenario and to remove all other factors. Games are nothing like that.

Do you understand what the test is doing?

Absolutely.

It is displaying how the AF implementation is working, and it does an extremely good job at it.

On a theoretical level maybe as the problems it displays do not exist in 99% of gaming situations. When I use my Radeon in actual games 99% of all surfaces are razor sharp as far as the eye can see, regardless of the polygon angle. I can see slightly less filtering in maybe 1% of situations if I stop playing properly and go specifically looking for them but really, that's just a dumb nitpick to even bring them up.

Me too. I spent my $200 on a R300 powered board and was shocked at how poor the texture filtering was.

I'm sorry, I don't know what else to say except you are very much the minority. Like I said before, the 9700 Pro with 16x performance looks better to me than the Ti4600 did at 8x standard. The textures are sharper, cleaner and crisper for much longer distances and the colour saturation and richness are fantastic. Also the texture rendering quality looks much better on the 9700 Pro while going back to NV2X boards makes feel like I'm dealing with plastic rendering as they lack realism and depth.

Listening to people all over they raved about how great it was.

With good reason I might add. You must be seeing something that no-one else sees in order for you to criticise it so heavily. And I can't for the life of me fathom what that could be as I don't see the problems you're referrring to.

Was there a configuration issue, or a driver problem previously with the Ti boards?

No, it's a known problem on Ti4xxx boards that they only utilise one texture unit when performing anisotropic filtering, thus heavily crippling performance. In fact my Ti4600 used to be slower than my Ti500 on occasion when both using 8x anisotropic.

Running Q3 16x12x32x8 without AA I hit 100FPS, far removed from what I would call a slideshow.

Yeah, and I used to get 108 FPS on my Ti4600 (IIRC). Now try those settings on Q3DM9 with yourself and all six bots in the outside area and watch the framerate hit an unplayable 30 FPS at times.

For RTCW, running a Ti4200 @stock speeds(mine is @290/560) 1280x1024 8xAF averages over 60FPS according to Digit-Life (Link).

Again it's just one demo. If you turn on 2x anisotropic it'll cut your framerate in half in the starting corridor outside of your cell. And you can forget about using 8x.

With noticeable mip banding? UnrealII, SS, SS:SE, Mafia, NOLF, NOLF2, Half-Life, CounterStrike, Blue-Shift, UT, Unreal, JKII and Sacrifice for mainstream games off the top of my head.

I've played/tested more than half of those games and I never spotted any mip-map transitions at all. In fact when I first got my Radeon 9700 Pro I was always checking the colormip levels in games just to make sure that I was indeed using performance mode as it looked unbelievably good.

Pretty much the only circumstance I didn't see noticeable mip banding is if I was playing games in which the AF was agressive enough to push the mip transitions back beyond visible range(Quake3 for example).

Uh, that's exactly the whole point. With 16x anisotropic the first mip-map boundary is pushed so far back that it's really irrelevant what happens after it. That's the beauty of Radeon chipsets: just leave it at 16x and forget about it. On the R300 the 16x setting produces razor sharp images as far as the eye can see, unlike pre-R300 boards which seem to drop to straight bilinear on the lowest mip-map level outside of the range of anisotropy. That is something that I don't like as it's clearly visible if the area you're looking at is quite large.

If you're griping about mip-map lines because you're using very low levels of anisotropy then you're completely going about it the wrong way. Either crank it up to 16x (which is what any sane person would do) or run quality anisotropic instead if they bother you so much, as that way you'll get normal trilinear filtering being applied outside of the region of anisotropy. And with a 9700 Pro/9800 Pro it'll probably still be fast enough to use 16x quality anistropic so you win both ways. Personally I don't see any real difference between the two so I use performance to get the whiplash speed that comes with it.

You can say the higher resolution is a way to solve anything,

It is. It'll solve anything except LOD/sharpness, which is exactly where anisotropic filtering comes in. Even very bad texture aliasing can be solved by the extra physical pixels that high resolutions provide. You think you've got bad texture aliasing? Then try 1600 x 1200 or 1792 x 1344 and get back to me.

Adjusting the LOD bias is a hack, you can do it on any board and you introduce rather severe texture aliasing adjusting it.

No, I mean move it back to 0.0 as you obviously don't like the Radeon's negative default. Personally I like it because I run at very high resolutions and they naturally counter the texture shimmering/sparkling that an aggressive LOD can produce.

No, transformation is not the main reason for texture aliasing nor is it directly associated with it(unless you are doing something horribly wrong).

I disagree with that comment as each level of resolution clearly reduces texture aliasing, along with many other things as well. And you don't need anisotropic filtering at all to achieve this.

Aliasing can be eliminated in terms or perceptible to a typical person running a resolution of ~8,000x8,000(upper limits of human perception for a typical viewing area) or a sampling that is comparable to such resolution.

Didn't I just say that high resolutions will cure texture aliasing?

Running ~64x anisotropic filtering on the NV2X style boards you should be able to pretty much eliminate mip mapping altogether, although obviously the performance hit would be staggering.

You might eliminate mip-mapping but you won't eliminate texture aliasing as long as your resolution is too low. If you run at 640 x 480 (for example) you can run whatever filtering modes you like and you'll still have problems, problems to which the only real cure is high resolutions. FSAA can help but it's not a true fix as it doesn't provde more physical pixels like high resolutions do. If you had an infinite resolution (or at least high enough so that the naked eye can't see the pixels) you could draw anything you liked and never have any artifacts since you'd essentially match the relative precision offered inside the video card.

By far the best method is to increase the anisotropic filtering and resolution as high as possible and the 9700/9800 boards really excel in this area. That is reality, not results from a colour wheel.

What settings were you running?

1600 x 1200 x 32 and in most cases even 2x anisotropic was unusable. And no, I never use FSAA or shadows.

Check the Digit-Life article I posted a link to above, the Ti4600 with max AF is faster then the 9500Pro with max AF running Quake3 1600x1200.

The Radeon is running 16x quality which is 128 tap sampling at the strongest level. The Ti4600 is only running 64 tap filtering. Also don't forget performance mode is available on the Radeon and if you use it you'll probably double the Ti4600's scores and it'll still look much better too.

And finally, the 9500 Pro is out-running the Ti4600 in the other anisotropic benchmarks, even with quality enabled.

 

[BenSkywalker]

Didn't you use to have a website?

Yes, kind of still do. Lost the server the site was on, currently working on getting it back up.

I know, from my own eyes running about 20 different games that the AF was better on my 8500 (this was including two flight sims) than both of my geforce 4s.

If you honestly think that, you are likely better off just ignoring AF all together and adjusting the LOD bias instead. You will eliminate the performance hit almost completely that way.

BFG

Yes but it's displaying them in a way that no game does. I've never seen any game that has a single, flat, high-contrast colour wheel in the middle of the screen. Have you?

It is taking a picture, top down, of the filtering implementation that a board is using(simplified). When you have a board reverting from true trilinear or not applying filtering in the proper form the wheel shows you where you will run in to problems. Non properly arcing mip boundaries, hard line transitions etc.

No, it's evident if every game looked exactly like the colour wheel did. The colour wheel is designed to bring out the absolute worst case scenario and to remove all other factors. Games are nothing like that.

It makes what the board is doing more pronounced for those that can't see it for themselves in game.

On a theoretical level maybe as the problems it displays do not exist in 99% of gaming situations. When I use my Radeon in actual games 99% of all surfaces are razor sharp as far as the eye can see, regardless of the polygon angle. I can see slightly less filtering in maybe 1% of situations if I stop playing properly and go specifically looking for them but really, that's just a dumb nitpick to even bring them up.

You are talking about LOD bias here, not AF.

No, I mean move it back to 0.0 as you obviously don't like the Radeon's negative default. Personally I like it because I run at very high resolutions and they naturally counter the texture shimmering/sparkling that an aggressive LOD can produce.

Running 1600x1200 the texture aliasing it still a neon billboard in the middle of an empty country road.

I disagree with that comment as each level of resolution clearly reduces texture aliasing, along with many other things as well. And you don't need anisotropic filtering at all to achieve this.

That has nothing at all to do with transformation as I stated.

Didn't I just say that high resolutions will cure texture aliasing?

Resolution 16 times higher then what you can run(assuming you can run 2,000x2,000).

You might eliminate mip-mapping but you won't eliminate texture aliasing as long as your resolution is too low. If you run at 640 x 480 (for example) you can run whatever filtering modes you like and you'll still have problems, problems to which the only real cure is high resolutions.

That is factually wrong. Watch FinalFantasy The Spirits Within on DVD. You do not need resolution to eliminate aliasing, it simply makes it simpler. As far as eliminating mip mapping, mip mapping is something you have to do right now(actually, some games not on the PC released in the last few years still don't do it). You have strong enough texture filtering and you eliminate the need for mip maps.

FSAA can help but it's not a true fix as it doesn't provde more physical pixels like high resolutions do.

You are talking about adding detail, that is not the same as reducing aliasing. Say you compare a monitor that is 15" running 800x600 against a 30" monitor running 1600x1200, which will have more noticeable aliasing all else equal? The answer it neither. Increasing the pixel density reduces aliasing as it increases the amount of samples taken eliminating some of the problem of sampling not being optimal for a given physical area. Running 4x AF you are sampling the same amount as you would if you quadrupled the resolution(8x6-16x12). What exact model of monitor do you have? In real life, I have yet to see a 21" CRT that could actually display 1600x1200. What 99.99% of people see at that setting, and even moreso for higher resolutions, is the approximation of the image based on the samples taken from the board. The same thing that sampling using a filtering implementation will get you.

The Radeon is running 16x quality which is 128 tap sampling at the strongest level. The Ti4600 is only running 64 tap filtering. Also don't forget performance mode is available on the Radeon and if you use it you'll probably double the Ti4600's scores and it'll still look much better too.

Performance mode is a non factor, it is bilinear. I understand that you can't see the mip transitions, but they are certainly there. Enable the colorized mip boundaries in SS if you don't believe me. Now saying for you bilinear filtering is OK is fine, Rollo may think the nV3X's Performance mode filtering is fine too. Dropping the R9500Pro to 8x doesn't have much of a performance impact IME. Dropping it to 8x would pull it about even to slightly ahead of the NV2X board. This is from the vastly superior anisotropic filtering on a performance basis?

And finally, the 9500 Pro is out-running the Ti4600 in the other anisotropic benchmarks, even with quality enabled.

True, in certain instances the R300 boards do show a decent performance boost running AF. Shader intensive ones in particular where the NV2X boards appear to have something out of whack fairly badly.

 

[oldfart]

New article on 5800 Vs 9800. Lots of AA/AF benches and screentshots..

 

[Manya]

I'd just like to add a few points. I own a 9700PRO and run everygame I have at 16x AF all the time. At no point have I ever seen a spot where it wasn't AF'ed properly. I suppose it's more because I don't sit there staring at a screenshot.

This debate is almost like the one about the FX5800 and the 16-bit FP vs ATI's FP 24-bit.

I don't think ben should beable to post a fair and judgmented opinion, because it's fairly obvious that he hates ATI and everything they stand for, and if someone gave him a 9800PRO he'd probably burn it and trade the ahses for a GF3...because they have better AF :/

 

[Rogozhin]

I would ask ben one question.

what gpu manufacturer pleases you most?

Rogo

 

[Rogozhin]

Ben

"If you honestly think that, you are likely better off just ignoring AF all together and adjusting the LOD bias instead. You will eliminate the performance hit almost completely that way."

chesnoya slova.


I already have the LOD adjusted as high as possible in both open gl and dx. Mipmap detail level is highest and TExture preference is also at the highest. I have AF set on 8x quality for both open gl and dx.

So what is it you are proposing again?

Rogo

 

[Pete]

I know wumpus has it in for ATi, but that doesn't mean Ben shared the same scarring childhood experiences. Seriously, Ben backs up his opinions far better than most, and you're free to disagree with them.

I think we should revisit the AF debate when nV releases the Det 50's and ATi releases the Cat 3.3's. So everyone hold yer horses for a few weeks.

 

[Rogozhin]

My eyes are more valid than anyone's review or review of a review.

rogo

 

[BFG10K]

It is taking a picture, top down, of the filtering implementation that a board is using(simplified). When you have a board reverting from true trilinear or not applying filtering in the proper form the wheel shows you where you will run in to problems. Non properly arcing mip boundaries, hard line transitions etc.

I know what the wheel is doing, the question is whether it's relevant to games, which it isn't. The scenario the wheel produces never happens in games.

It makes what the board is doing more pronounced for those that can't see it for themselves in game.

Well that just proves my point doesn't it? You need the wheel to see problems because you can't see them in games. That doesn't make the wheel terribly useful.

You are talking about LOD bias here, not AF.

No, I'm talking about the non-filtering that colour wheel shows. If you can't see them in games then what good is the colour wheel other than for theory and heated debate?

That has nothing at all to do with transformation as I stated.

The fact is that you have essentially an unlimited resolution and you're trying to put it into a limited resolution. A lack of physical pixels always produces a number of artifacts and each resolution bump decreases them. If you don't believe me then try it for yourself.

Resolution 16 times higher then what you can run(assuming you can run 2,000x2,000).

I didn't say 1600 x 1200 was the pinnacle of resolutions, all I said was that it's pretty good at removing the texture aliasing and that each step higher reduces it even further.

That is factually wrong. Watch FinalFantasy The Spirits Within on DVD. You do not need resolution to eliminate aliasing, it simply makes it simpler.

Hang on, it sounds like you're now talking about edge/polygon aliasing rather than texture aliasing. Yeah, for edge aliasing you can certainly use any form of FSAA to clean it up for you, just don't tell me that movies don't have any edge aliasing because they sure as heck do. Of course you might be watching the DVD on a HDTV resolution in which case that just proves my point yet again.

As far as eliminating mip mapping, mip mapping is something you have to do right

I didn't even mention mip-mapping with respect to resolution. Resolution will do nothing with it, which is where anisotropic filtering comes in.

Say you compare a monitor that is 15" running 800x600 against a 30" monitor running 1600x1200, which will have more noticeable aliasing all else equal?

No, I'm saying to run the same monitor at both 800 x 600 and at 1600 x 1200 and you'll have much less aliasing at 1600, both edge and texture.

Performance mode is a non factor, it is bilinear.

It certainly is a factor because you get quality level image quality at only a fraction of a performance hit. This is why Radeons are so popular.

I understand that you can't see the mip transitions, but they are certainly there. Enable the colorized mip boundaries in SS if you don't believe me.

Of course they're there and I'd be an utter looney to deny that a bilinear filtering mode is not, uh, sampling bilinear texels. What I'm saying is that without the colormips you can't see evidence of the bilinear filtering.

This is from the vastly superior anisotropic filtering on a performance basis?

Yeah it is superior. If you benched the Radeon at 16x performance it would both eat the GF4 alive and look better too.

 

[Rogozhin]

Good response BFG

Rogo

 

[BenSkywalker]

Manya-

I don't think ben should beable to post a fair and judgmented opinion, because it's fairly obvious that he hates ATI and everything they stand for

I paid my money to purchase an ATi product. I read the reviews and listened to people's comments in forums. Then I actually got the product and it was considerably different then what people claimed. The AA was right where people said it was, only I have never been a big fan of AA(I've been here for years, anyone can check my post history on the subject). I am and have long been a proponent of proper texture filtering implementations, strong T&L units and solid drivers. I was lead to believe that ATi now posessed all of those things in stellar fashion. They have the T&L unit down pretty well, they are lacking in the others. Notice me slamming ATi's AA? Nope, and you won't. Do you see me ripping on their speed? No again, and you will not. Do you see me blasting them for missing features(and there are some already in use they are missing)? No again. I take issue with the things that people stated ATi was doing extremely well that they are certainly not.

Pete-

I know wumpus has it in for ATi, but that doesn't mean Ben shared the same scarring childhood experiences.

Hehe, Jeff had it in for ATi because he was fooled by the hype around the original Radeon. His distaste for ATi truly paled in comparison for his loathing of 3dfx though

Rogo-

My eyes are more valid than anyone's review or review of a review.

For you. I'm talking about doing things the right way versus not, and my eyes did see the mistakes quite clearly. Inaccurate rendering has long been another pet peeve of mine. I have not been too kind for the FX's AF implementation if anyone really wants to notice, I just don't seem to have people as upset by that fact.

I already have the LOD adjusted as high as possible in both open gl and dx. Mipmap detail level is highest and TExture preference is also at the highest. I have AF set on 8x quality for both open gl and dx.

So what is it you are proposing again?

Have to spell it out? Go in to the registry and isolate out the values which pertain to LOD bias and adjust them beyond that which is allowed by the driver's default control panel.

BFG-

Well that just proves my point doesn't it? You need the wheel to see problems because you can't see them in games.

No, more accurately the wheel exists because you can't see them in games. I can, that is my problem.

The fact is that you have essentially an unlimited resolution and you're trying to put it into a limited resolution.

This is not true, not unless you are talking about procedural based textures the likes of which have not been seen. You have a very clear finite resolution for textures utilized in games.

Hang on, it sounds like you're now talking about edge/polygon aliasing rather than texture aliasing.

Absolutely not, I'm talking about texture aliasing. Watch FF:TSW and try and spot texture aliasing.

It certainly is a factor because you get quality level image quality at only a fraction of a performance hit.

If you can't see the mip banding then perhaps the IQ is equal for you, I can see them quite clearly and it is certainly not.

Yeah it is superior. If you benched the Radeon at 16x performance it would both eat the GF4 alive and look better too.

So run bilinear on the GF4 and adjust the LOD bias to match what the Radeon uses.

 

[Rogozhin]

Ben

I never proposed that you weren't seeing problems, that was not the implication of my statement.

"For you. I'm talking about doing things the right way versus not, and my eyes did see the mistakes quite clearly. Inaccurate rendering has long been another pet peeve of mine. I have not been too kind for the FX's AF implementation if anyone really wants to notice, I just don't seem to have people as upset by that fact."

Besides the source you pulled on digilife, in what games did you see rendering errors, were you able to get screenshots (if it's possible to see the problem in stills), and can you post them if you have them?

Rogo

 

[Rogozhin]

Hm

I went into Local machine into software into ati tech inc, into drivers and there is no reg key string, no lod, nothing.

Where is it hiding?

Rogo

 

[Rogozhin]

Ben

One more thing.

If you are saying that ATI's rendering method is cutting corners in its texture ops on near-field objects then you will have to give empirical proof. Because from I've read the rendering methods on ANY video card, in their optimizations eliminating non-visible geometry or pixels (culling, clipping, depth-buffering, backface culling, etc.), or reducing texture and geometric resolution using a distanced-based algorithm (MIP-mapping, geometric LOD), non of these methods reduce overall IQ since texture and geometric detail are rendered where they are needed most (rather close to the pov).

Also here is a quote from David Kirk regarding Nvidia's adaptive method of AF.

"The goal here is to correctly sample the footprint of the source texture that lies within the pixel, including orientation, and perspective. What anisotropic filtering does better than simple trilinear is those last two: orientation and perspective.
Anisotropic is not necessarily a rectangle, since the direction of anisotropy may not be axis aligned. Also, the actual shape may be more of a quadrilateral, due to perspective. The combination of these is very complex - the nearer edges of the quadrilateral will be larger, and thus lower LOD (level of detail) and weighted more. Also, because of the perspective effect, the entire "front half" of the sample is often weighted more than twice as much as the other half.

Consequently, the LOD may change throughout the pixel, by a lot more than 2x

Our anisotropic is a weighted average of up to 8 trilinear samples, along the major axis of anisotropy. So, we may include up to 64 samples, but the samples may be taken from only two MIP maps, and the weighting is nonlinear. The samples are effectively closer together in the part of the pixel that is nearer to the eye.
The goal here is to correctly sample the footprint of the source texture that lies within the pixel, including orientation, and perspective. What anisotropic filtering does better than simple trilinear is those last two: orientation and perspective. "

What I take from this quote is his saying that they use an adaptive method just as ati does the only difference being that the "footprin" is shaped differently, but still on the same z axis, Meaning both cards render only those parts of the screen that can be seen.

Any discussion appreciated

Rogo

 

[chsh1ca]

Yes but it's displaying them in a way that no game does. I've never seen any game that has a single, flat, high-contrast colour wheel in the middle of the screen. Have you?

I guess you've never played Super Single Flat High-Contrast Colour Wheel Floating In The Middle Of The Screen then eh?

 

[BenSkywalker]

Besides the source you pulled on digilife, in what games did you see rendering errors, were you able to get screenshots (if it's possible to see the problem in stills), and can you post them if you have them?

You'll only see the rendering errors in motion. The mip banding was noticeable in every game I checked that had a reasonably large viewing area(almost everything outside of Quake3 style confined games). The improper levels of AF is apparent in numerous instances, although they are even easier to miss then the mip map transitions. Besides Digit-Life, the B3D article covers it and there have been numerous discussion over there about it.

Because from I've read the rendering methods on ANY video card, in their optimizations eliminating non-visible geometry or pixels (culling, clipping, depth-buffering, backface culling, etc.), or reducing texture and geometric resolution using a distanced-based algorithm (MIP-mapping, geometric LOD), non of these methods reduce overall IQ since texture and geometric detail are rendered where they are needed most (rather close to the pov).

Anything that is done with surfaces not visible is fair game, the PVR based boards don't render anything that isn't visible. The problem is when you start to impact visible objects. Mip mapping reduces texture clarity, it is a hack that had to be used to eliminate serious texture aliasing and to date we still don't have a filtering technique strong enough to eliminate it. Yes, I do consider mip mapping a hack of sorts, when we have a good enough texture filtering implementation we will no longer need it. Using the latest boards it is already fairly uselss under certain titles(Quake3 for example) when running max AF. Geometric LOD is a hack to reduce the poly complexity of models, if it has noticeable transitions between tesselation stages then I do have an issue with it. One of my examples for a clear edge that hard T&L used to have over software was Sacrifice which uses a geometric LOD system, with hard T&L the tesselation changes were far less frequent.

What I take from this quote is his saying that they use an adaptive method just as ati does the only difference being that the "footprin" is shaped differently, but still on the same z axis, Meaning both cards render only those parts of the screen that can be seen.

Anisotropic filtering, done properly, is adaptive. When I talk about adaptive AF implementations I'm talking about those that cut corners and aren't sampling all things on an equal basis(which does not mean all things should be sampled the same). The FX boards are also adaptive implementations, and while they aren't quite as out of whack as the R300 core boards in terms of sampling, they are not directly comparable to the NV2X line of boards which do it the right way.

I went into Local machine into software into ati tech inc, into drivers and there is no reg key string, no lod, nothing.

Where is it hiding?

Run a search through your registry(Ctrl+F- use F3 to find the next) searching through all entries listed as ATi. There are likely ten or twelve different locations for ATi throughout your registry located in different spots. One of the entries will adjust LOD bias, although it may not explicitly state LOD bias. A few years back nVidia used to 'hide' their OCing controls by using the Simpsons character names for their clock rate settings. Likely the people over at Rage3D will know which registry value adjusts the LOD bias and where it is located(it depends on which OS you are running where it is, and what the valid values are for entry).

 

[Rogozhin]

Ben

Thank you for the directions to find LOD bias adjustments.

In the comparisions I've read and seen the mipmap banding (mipmap boundry conspicuousness?) plus in game, ati's trilinear AF has less demarcation of mipmap banding than both the nv30 and nv2x. Anand, Hardocp, and Tom's all showed this in their most recent reviews.

And I will ask once more what specific games and in what specific level, eniorments, with what specific settings on the card do you see rendering errors? The quote below does not answer my question. And since I already said that it will probably not work to take screenshots to show these errors (which you once again mentioned AFTER i said this), could you send me or post on you webhost an ingame video of these errors, OR, take a video capture of the problem, burn it and mail it to me, i will pay for CD and shipping.

"You'll only see the rendering errors in motion. The mip banding was noticeable in every game I checked that had a reasonably large viewing area(almost everything outside of Quake3 style confined games). The improper levels of AF is apparent in numerous instances, although they are even easier to miss then the mip map transitions. Besides Digit-Life, the B3D article covers it and there have been numerous discussion over there about it."

Rogo

 

[BenSkywalker]

And I will ask once more what specific games and in what specific level, eniorments, with what specific settings on the card do you see rendering errors?

Mafia, Sacrifice, Serisous Sam, SS:SE, NOLF, NOLF2, Unreal2, UT and Unreal for games to start with(it is apparent in pretty much every game, though not quite). Settings use AF, any setting, performance or quality(it doesn't matter). I'm not talking about missing textures or dropping entire polys, I'm talking about improper texture filtering resulting in artifacts. Well, there are the rendering errors in Sacrifice and Quake3 engined games, those can be shown in screenshots. Give me a sec, here is a shot of the Sac rendering errors-

Sac Z fighting.

And since I already said that it will probably not work to take screenshots to show these errors (which you once again mentioned AFTER i said this), could you send me or post on you webhost an ingame video of these errors, OR, take a video capture of the problem, burn it and mail it to me, i will pay for CD and shipping.

You can't see the issues with the R200 core boards which are absolutely horrible, there is no chance you will notice it on a R300 core board.

 

[Rogozhin]

Ben

The black triangles around the doorway?

I have Mafia and UT2003 (many others but those are some you listed).

If you can point me towards what to look for in either game I will be happy to look for the errors.

I did this with Chalnoth on nvnews.net because he was saying there were z-buffer errors in morrowind. So he told me where to look on what map in an exact location and BINGO z-buffer errors, water shifting height and small squares on certain guards' armor.

If you can direct me the same way I would appreciate it.

Rogo

 

[Rogozhin]

There are some good af discussions going on at nvnews, along with some benchmarks with people that own the FX 5800 ultra, go check it out fellas

rogo

 

[BFG10K]

No, more accurately the wheel exists because you can't see them in games. I can, that is my problem.

I understand that you can see them, all I'm saying is that most people can't and thus using the colour wheel to convince them isn't any good. I can't see weak filtering unless I stop playing the game and specifically look for bizarrely angled surfaces and I might find the odd rare one that has problems. But even if I stop and look for the mip-map lines I can't see them unless I use the colormips.

What I do know is that if I play the game normally and don't try to look at it with a microscope then the image quality looks awesome. The same can't be said for pre-R300 anisotropic filtering though and I fully agree with your findings.

Absolutely not, I'm talking about texture aliasing. Watch FF:TSW and try and spot texture aliasing.

Texture aliasing can be cured by a negative LOD and/or any form of blurring. That doesn't mean that it's the best way to do it.

So run bilinear on the GF4 and adjust the LOD bias to match what the Radeon uses.

How is that going to help if the Radeon's 16x performance already looks better than the Ti4600's 8x full? Going to bilinear on the Ti4600 will only look worse since you'll be sampling half the texels that the Radeon is sampling. Also the Ti4600 will still yield slideshow performances.

 

[BenSkywalker]

How is that going to help if the Radeon's 16x performance already looks better than the Ti4600's 8x full? Going to bilinear on the Ti4600 will only look worse since you'll be sampling half the texels that the Radeon is sampling. Also the Ti4600 will still yield slideshow performances.

It will improve performanc considerably and it will also give roughly the same sharpness as ATi's 16x Performance AF. If texture aliasing isn't an issue for people, then using a negative LOD bias adjustment will net you the best results with the lowest performance hit.

But even if I stop and look for the mip-map lines I can't see them unless I use the colormips.

Don't stop, that actually makes them very hard to pick out Just move straight forward and look back a bit from where they would be using a non AF based implementation.

Texture aliasing can be cured by a negative LOD and/or any form of blurring. That doesn't mean that it's the best way to do it.

With a render engine like the one used in FF:TSW, you are talking about thousands of samples per pixel. You filter something enough, it does eliminate noticeable aliasing.

Rogo-

I have Mafia and UT2003 (many others but those are some you listed).

If you can point me towards what to look for in either game I will be happy to look for the errors.

Improper texture filtering is what I'm talking about. This is much harder to spot then Z fighting, which is one of the worst types of image corruption. For texture filtering issues the main two I'm talking about is running performance mode AF which results in bilinear filtering being used(actually, for your board it doesn't matter what setting you use). There will be a line where textures transition from sharper to less sharp, and that line is fairly clearly defined. The other end of improper texture filtering is texture aliasing and the occasional texture which appears out of whack as it is not being filtered the same as those around it. This may be a texture too sharp so it is 'shimmering' while in motion or it could be one that is too blurry compared to its surroundings. This is not just an issue with ATi boards, the NV30 core boards also have less then perfect AF at the moment in terms of implementation and performance(ATi only has issues with implementation, the NV2X boards only have issues with performance).

 

[BFG10K]

It will improve performanc considerably and it will also give roughly the same sharpness as ATi's 16x Performance AF.

Really? I've used bilinear at every filtering stage using RivaTuner and it didn't seem to help performance much at all.

Also nVidia's 8x bilinear is only 32 tap instead of ATi's 16x performance being 64 tap.

Don't stop, that actually makes them very hard to pick out

Stop as in "stop playing the game and look for artifacts", not flat-out stop.

Just move straight forward and look back a bit from where they would be using a non AF based implementation

A technique I like to use that sometimes works even better is to repeatedly tap the forward key. If they're present you can see the formation of the mip-lines if you pause right in between when they start to form.

(And yes, I have tried both proper movement and my technique and the 9700 Pro passes with flying colours.)

With a render engine like the one used in FF:TSW, you are talking about thousands of samples per pixel.

But thousands of samples per second are not feasable for realtime rendering like games. High resolution is though.

 

[Rogozhin]

Ben

So this "actually, for your board it doesn't matter what setting you use," means that you believe that the errors occur even when running trilinear AF?

Rogo

 

[BFG10K]

No, what he's saying is that non-R300 chipsets do not support trilinear anisotropic, only bilinear. That means that the only option you have is performance mode.