Explanation of ATIs AF advantage?

[nRollo]

It does less, so why shouldn't it be faster?

"At any angle the NV30 shows the same picture which depends only on the plane's inclination. The R300 copes only with the angles of 0 and 90 degrees and angles close to 45. At all intermediate angles (20, 30, 60, 70 etc.) the ATI's algorithm works much worse."

"At a low anisotropy degree the ATI's algorithm behaves similarly to the NVIDIA's one selecting MIP levels correctly according to the real distance (an ideal picture must represent circles), but at a higher anisotropy level the NVIDIA's and ATI's methods use different approaches. I must say that the NV30 works more correctly in case of longer distances (far tunnel end) and a high anisotropy level. "

Digit Life's review doesn't seem to agree with the popular opinion around here that ATIs AF implementation of AF is "superior" and yields "better IQ".

"First of all, I want to destroy the myth that ATI uses a variation of the RIP mapping and additional narrowed texture versions. Nothing of the kind - the anisotropic filtering works with the same source textures as the NVIDIA's one does, and no extra space or extra calculations are used. They have entirely different filtering algorithms. Firstly, the samples are selected one by one, without preliminary bilinear filtering. The filtering is indirect here - the algorithm decides from what MIP level it should take a sample; on the picture above small squares refer to samples from a more detailed MIP level, and bigger squares are from a rougher one. So, we can take much more samples using same computational resources and bandwidths, but the samples won't be originally filtered which may result in inferior quality."

BTW- check out the DOOM3 benches- at those framerates, no one will be running FSAA/AF, but look at how the 9700 is smacked down....

Do I care about 15fps at RTCW? No. Do I care about 15fps 39 vs 54 DOOM3? Err, yes.

 

[Rogozhin]

I've read this article.

One of the main retorts to this is that almost no fps or any other games besides flight simulation use angles other than 90, 45 and 0.

When i first hear about this AF implmentation by ati (at the time I was running a 8500, msi geforce 4 ti4400, and gainwaward geforce 4 ti4200) I tried out full af on all cards on IL-2 sturmovik, Ghost Recon, and Dungeon Siege. I never did notice any form of degredation when compared to the geforce 4 cards and since the fx uses the same implementation of AF as the geforce 4s I assume that this conclusion of mine still holds true. If i had the money to purchase a FX I would do it and compare it so I could say definitively.

I also posted a thread about this over at nvnews.net a few months ago but I'm to lazy to find the url sorry rollo.

Rogo

 

[BoberFett]

After Nvidia got beat down in 3dMark03, Nvidia fanboys broke out in cries of "But do you sit around and play 3dMark all day? It's about the games!" (To say nothing of the fact that they loved it when Nvidia was on top of the 3dMark world)

So I'll just ask "Do you sit around watching textured tunnels all day or do you play games?"

The general consensus is that in games (not tunnel texture demos) the ATI card produces better image quality.

 

[nRollo]

"So I'll just ask "Do you sit around watching textured tunnels all day or do you play games?"

Watching texture tunnels is key to me.....

I play games obviously, but it's easier for me to tell which card has more accurate rendering from tests like this than screenshots of games on a website.

 

[BenSkywalker]

The general consensus is that in games (not tunnel texture demos) the ATI card produces better image quality.

Versus the FX using the lower AF settings it appears that way, not vs the NV2X boards(although the R300 core boards are much faster). A lot of the edge that sites bring up is a mixture of LOD bias and color saturation, not AF quality. Running the highest quality AF setting the FX appears to be superior to the R300 boars, although performance is considerably lower.

One of the main retorts to this is that almost no fps or any other games besides flight simulation use angles other than 90, 45 and 0.

That's not true. If you are facing a wall and turn in a circle you force textures in to all sorts of odd angles compared to the POV. It doesn't matter what the native orientation for the texture is, proper anisotropic is calculated based on the POV, not the world space alone. Even ignoring that, current games are using considerably more intricate geometry then in years past offering considerable amounts of varrying angles even when looking just at world space.

When i first hear about this AF implmentation by ati (at the time I was running a 8500, msi geforce 4 ti4400, and gainwaward geforce 4 ti4200) I tried out full af on all cards on IL-2 sturmovik, Ghost Recon, and Dungeon Siege. I never did notice any form of degredation when compared to the geforce 4 cards and since the fx uses the same implementation of AF as the geforce 4s I assume that this conclusion of mine still holds true.

The R300 uses a different AF implementation then the R200 and the NV3x uses a different implementation from the NV2x. The R200's is hands down the poorest I've ever seen. Want an example- crank the AF to the highest setting and fire up a flight sim and do a roll(massive texture aliasing with the R200 core boards).

 

[CurtCold]

I think benching the Alpha version of Doom3 is 1st off unprofessional, and second not actually showing the true performance that will be when Doom3 is actually released.

For the rest of the AA, and AF statements, each company has to choose which methods they prefer, and how to better benifit the performance of their respective cards. I was under the impression that ATI's main advantage with AA, and AF was the 256Bit memory bus interface.

 

[nRollo]

Curtcold:
"I think benching the Alpha version of Doom3 is 1st off unprofessional, and second not actually showing the true performance that will be when Doom3 is actually released."
I'm not against them using the Doom3 alpha to benchmark, unless I buy a 5800 based on this and ATI comes out on top in the final release. I doubt it though. Carmack said in his plan he's using a 5800 Ultra to code Doom 3 because he ran into limitations with the R9700 Pro. Also, Carmack has been high on nVidia, not ATI for some time now. If I remember right, he usually says on paper their chips are better, but in practice they're not.

"For the rest of the AA, and AF statements, each company has to choose which methods they prefer, and how to better benifit the performance of their respective cards."
Like Ben Skywalker said, you can't get a feel for accuracy of AF looking at still screenshots. I think you'd have to see the video of panning around curved objects to see the difference from what I can gather from the digit life article.

 

[CurtCold]

I do remeber seeing the Carmack comments about using the Nvidia card. I think his reasoning was not only performance, but better drivers by Nvidia.

Also remeber previous reviews of ATI vs. Nvidia AA, and AF. I remember reading that ATI's AA and AF was not on the same level of Nvidia's. Even though, I thought through recent reviews that ATI had improved upon previous AA, and AF performance. Maybe I'm wrong.

Nevertheless there are currently no perfect product out, and probably won't be. There will be issues with ATi's drivers, and FX fan noise etc, for years to come.

 

[SickBeast]

Rollo do you have your $400 hairdryer yet? And why are you comparing the 9700 instead of the 9800 with the gffx? And, I didn't see anything about them fixing the driver settings in that review so that it is fair. If you look at the hardocp reviews, they found that the gffx drivers have lower image quality at the same settings as the ati cards. Therefore, the review you linked to isn't making a fair comparison.

And as for the AF thing, EVERY other review that I've read has stated that the ati cards have better image quality. I'm pretty sure there is an option in the ati drivers to allow "quality" AF, which should be the equivalent of what the nvidia cards do.

 

[J5im8yo]

Camrack is biased, he blames ATi for the leak of DOOMIII alpha

 

[kylebisme]

Originally posted by: RolloBTW- check out the DOOM3 benches- at those framerates, no one will be running FSAA/AF, but look at how the 9700 is smacked down....

Do I care about 15fps at RTCW? No. Do I care about 15fps 39 vs 54 DOOM3? Err, yes.

lol ya check that they are showing performace on illegally obtained software which is no where near its final state and with performance figures much different than recent accounts from the man behind the game himself. that does a lot to show how creditable the site is. not to mention that anyone who looks at the screenshots they present as well as ones from some other reviews make it obvious that xbit is blowing the situation way out of proportion.

 

[Rogozhin]

Ben

"Want an example- crank the AF to the highest setting and fire up a flight sim and do a roll(massive texture aliasing with the R200 core boards). "

All I play are flight sims. That's exactly what I did with my geforce 4s and 8500, I turned on 16x af on the radeon and the highest af option on the geforce 4s, can't remember the setting since i was using rivatuner and that was 8 months ago. The 8500 had much better af than the geforce 4s, cockpit textures were much crisper (this was also the case with nascar 2002).

Texture aliasing is what now? Doesn't aliasing cover everything? I've never heard the term, please explain.


And you are saying that the AF when maxed on both the fx and radeon 9800 the fx looks better?

That is taken from what source?

Surely not anand or Hardocp.

Rogo

 

[BenSkywalker]

The 8500 had much better af than the geforce 4s, cockpit textures were much crisper

How crisp the textures are can be adjusted through the LOD bias. If you use an agressive enough LOD bias setting you can make any board, even those with no AF, have textures as crisp as the R9800Pro and moreso. The problem is if you are too agressive with the LOD bias you introduce massive texture aliasing.

Texture aliasing is what now? Doesn't aliasing cover everything?

Texture aliasing can best be described as 'shimmering' textures. It gives the appearance of moire patterns that shimmer considerable while in motion. Taking screen shots it is pretty much impossible to display texture aliasing, you must see it in motion. What anisotropic filtering is supposed to allow is an increase in the sharpness of textures without introducing texture aliasing. What ATi appears to do is adjust the LOD bias for their optimal angle, which unfortunately has a noticeable superior filtering edge when compared to their worst case angles. What you end up with is very noticeable texture aliasing on less then optimal angle on the R300 core boards.

The R200 core boards(which includes the R8500) have a problem with anisotropic filtering when rotating on the Z axis, they don't utlize any. The LOD bias remains adjusted as if the board were filtering on its optimal angle but no AF at all is being applied. This causes massive texture aliasing.

And you are saying that the AF when maxed on both the fx and radeon 9800 the fx looks better?

Which looks overall better will be a matter of personal preference. Texture aliasing doesn't bother some people as much as others. What I'm saying is that looking at the actual AF implementation the FX's highest end anisotropic filtering is a better implementation in terms of quality. It is, however, much slower then the R300's highest setting. The NV2X boards have the highest quality AF implementation that I have seen. The R300 and NV30 boards all utilize an adaptive implementation that cuts corners, the FX cuts more corners in its lower performing modes while cutting less using its highest setting. The NV2X boards do actual full anisotropic filtering and they offer the best solution looking at the actual filtering technique.

Some of the sharpness of the ATi boards can also be attributed to the amount of color saturation that they utilize when compared to the nV boards. This makes for a brighter appearing image however it also introduces color bleeding. nVidia boards can be easily adjusted to give the same appearance, as ATi boards can be calibrated to look like nV boards(for nVidia utilize the digital vibrance settings and crank them up).

That is taken from what source?

The Digit-Life article is the best I've seen in breaking down exactly what the boards are doing. I was running a Radeon9500Pro for about a month also and have had plenty of time to witness the problems with the implementation.

Surely not anand or Hardocp.

Anand stated that there was virtually no difference between ATi's Performance and Quality settings, while one of them utilizes bilinear filtering and the other utilizes trilinear. I see it as d@mn near night and day, easily noticeable mip banding using Performance settings while none using Quality. I picked up a R9500Pro in no small part based on reviews such as Anand's, my own experience showed considerably different characteristics then what Anand discussed. As far as Kyle goes, he is now using an ATi shader demo for one of his benchmarks to compare ATi and nVidia and presenting it as fair. I'll take B3D and Digit-Life for my vid card reviews.

In terms of AF, the NV2X boards are the best but are very slow. When looking at the R300 v the NV30 the NV30 boards offer the best implementation but it also considerably slower then that offered on the R300. The modes which run at comparable speed to the R300 lack the overall IQ of the R300 parts. Most people will factor speed against IQ and that is where most of them will come to the conclusion that ATi's parts have the best overall IQ.

Where the R300 boards really shine is in their non SS AA implementation. With the exception of older titles that utilize alpha textures heavily the R300 core boards do have rather vastly superior AA at 4x and particular 6x. ATi could stand to offer a super sampling implementation to utilize on older titles(such as the HL games), but on anythting reasonably current, or when looking at overall edge AA ATi's 6x or 4x AA is easily superior to the implementation in the NV3x or NV2x boards.

 

[Rogozhin]

Ben

I don't know enough about AF to understand what you're talking about would you mind if I posted your expanation on nvnews.net and rage3d.com?

Rogo

 

[Rogozhin]

Ben

The "shimmering textures" are seen on 'what' in a flight sim like IL-2 or FB?

There's never been a post about this over at il2sturmovik.com in the tech section, would you mind posting your findings over there to see what our resident 3d gurus have to say?

thanks again!

rogo

 

[nRollo]

I like how everyone throws Kyle/HardOCP out there as a reliable reviewer.
He's entertaining, but not exactly what I'd call an authority. I'd say Ben Skywalker is a more relevant source of info than Kyle. At least he seems to understand the issue at hand.

 

[Rogozhin]

you mean ben is a more "reliable" source?

And how so?

Rogo

 

[kylebisme]

Originally posted by: Rollo
I like how everyone throws Kyle/HardOCP out there as a reliable reviewer.
He's entertaining, but not exactly what I'd call an authority. I'd say Ben Skywalker is a more relevant source of info than Kyle. At least he seems to understand the issue at hand.

you mean at least Ben is backing up the argument you presented? and you don't make yourself out to be a very creditable source of info it has been quite some time sense Kyle has done a videocard review anyway.

 

[Pete]

Damn, Ben, that's a quality explanation. Didn't expect you to put so much effort into it--might as well save that for a quick copy 'n paste for the inevitable repeat Q. (Speaking of effort, I miss GameBasement. )

Both cards take shortcuts with AF (though the FX allows you to use the highest-overall quality AF in Application mode, albeit at a significant framerate hit). In the Quality mode, ATi doesn't filter every angle, and nVidia doesn't do true trilinear. It's a trade-off as to which method is more important to the games you play.

I'd like to see a detailed AF comparo once nVidia releases the Det 50's and ATi releases the Cat 3.3's, with as wide a variety of games (read: not all FPS's) as this review: http://www.3dvelocity.com/reviews/gffx5800u/gffx_5.htm.

 

[nRollo]

"And how so?"
Ben's explanation of why he feels ATIs AF is lacking is better than 99% of the reviews I've seen. He gives examples from his experiences, then explains why he thinks the anomalies are happening, referencing respective LOD bias settings, color saturation, and the AF methods employed by each card. Compare that to this site's "Here's some screenshots, it looks better to us" approach.

"you mean at least Ben is backing up the argument you presented? and you don't make yourself out to be a very creditable source of info "
Yes, and I'm no Ben or Digit Life guy. I configure and demonstrate the use of software for a living, not review VGAs. (and I haven't had the time to learn Ben's level of expertise in the mechanics of AF methods lately- 3 year old keeps me running)

 

[oldfart]

Another review comparing AA/AF

 

[BenSkywalker]

Rogo-

There's never been a post about this over at il2sturmovik.com in the tech section, would you mind posting your findings over there to see what our resident 3d gurus have to say?

Texture shimmering isn't a measurable issue, at least not that I am aware of. If you have a V5 handy I can tell you how to see texture aliasing easily, go in to the driver panel and set the LOD bias to -3, then fire up any game. That will demonstrate to you what texture aliasing is. As far as 'my findings' on the problem with the R200's AF implementation, it is something that is fairly heavily documented. IIRC many reviews at the launch of the R300 core boards mentioned that this serious issue was no longer present. As far as how it impacts that particular game, I can't speak intelligibly on the subject. The sharper the textures are to start with, the worse it will be when a too agressive LOD is used. There are actually several older titles that utilized very low res textures that can stand to use a fairly agressive LOD bias adjustment without any real ill effect. Stating that the problem exists for certain in a particular game I don't have hands on time with isn't something I would do. I can vouch for the issue on numerous others though(and there are plenty of people that can speak on the problem with the R200 core boards and their Z axis/AF bug).

I don't know enough about AF to understand what you're talking about would you mind if I posted your expanation on nvnews.net and rage3d.com?

Nothing I've stated is a big secret. Pretty much every bit of it is covered in the Digit-Life article, and backed up by some of those that were posted later(by Pete and OldFart). The big problem with the FX's AF is performance when comparing the highest quality settings for both. When comparing the lower quality settings the FX gets spanked in terms of implementation. Both the R3X0 and NV3X are inferior to the NV2X when it comes to AF quality, but both of them slaughter it in terms of speed. Most people consider AA and AF a tradeoff between speed and performance, and in that aspect right now the NV30 looks quite poor in comparison to the R3x0 core boards. That doesn't change the quality of it at the highest settings. Almost all the reviews focused on LOD bias or color settings when they compare the AF implementation, the Digit-Life review gives a complete breakdown of exactly what is going on(as long as you know the basics it is quite apparent). As far as posting at nVnews, nothing I have to say would be breaking news to any of the guys that write for the site(they post @B3D), the problem most people have is performance. I won't post @Rage3D. I attempted to go over there once to help out some people getting the Vogel S3TC patch working for UT and was flamed quite quickly because I was running a nVidia board at the time and I had made comments that the Radeon wasn't the be all end all on another site(here, IIRC). Tried to defend myself and had my posts deleted by a mod, including my initial post where all I did was ask people for what settings they were running. After the fact the mod that did so apologized(about erasing my post that was simply trying to help people), but he couldn't be bothered to read the contents of my posts prior to erasing them. That was about two and a half, three years ago IIRC. I may try and post there again in another couple of years to see if they have chilled out.

The Snowman

and you don't make yourself out to be a very creditable source of info it has been quite some time sense Kyle has done a videocard review anyway.

I don't know about Rollo, but I refer to HardOCP as 'Kyle's' or 'Kyle' all the time. Same as I do with Anand's, even if Anand himself didn't do the review. When I was doing reviews, the site I wrote for always got the credit, and Kyle is a lot quicker to type then [ H ] ard OCP

On the broader subject of vid card reviews at Kyle's, there was a lengthy discussion about his review methodology back in late '99 early 2K here when Kyle was trying to demonstrate the viability of a new feature on a graphics board and totally blew it(trying to demonstrate T&L's ability but he ran a fill limited situation). This in itself wasn't too big of an issue however when he was informed of how to display more accurate results for what he was trying to do(which only required that he lower the resolution) he refused to do so and made no attempt to ever rectify the misconception that he had given a lot of people(IIRC it was myself and LeoV that really went off about it). I'll look at the benchmark results from their reviews, but I don't but much faith in the commentary. Yes, they are much better now that Kyle is not doing them himself, but they still leave much to be desired(as do most sites).

Pete-

Speaking of effort, I miss GameBasement.

Thanks I'm working on getting it back up right now. Found a host, working on redesigning the whole site and bringing all of the old content over to the new format. Also working on getting a decent forum going for it too. Progress is a bit slower then I was hoping, but hopefully it will be up and running before too long

 

[Rogozhin]

Another choice when comparing the AF on the cards (since nvidia has implemented new names for it's settings) is to run the highest AF available for both cards (and not what nvidia recomends, this is explained in the article of Kyle's).

This is WHY he chose to run the AF settings 8x quality on the radeons and 8x application (which is nvidia's highest AF setting). And don't forget that the radeon can do 16x quality if you really want the high IQ.

And if the LOD bias IS adjusted through ati drivers to enchance the look of its AF why would that be a bad thing. I still see no proof of ati doing that. If he wants to back up his proposition then he better give us some videos, or at least some screens.

I would LOVE to own an FX just to compare it to my radeon, alas, the tax refund isn't big enough and I can't take that much out of the checking account without my wife knowing (like rollo's here).

Rogo

 

[kylebisme]

Originally posted by: BenSkywalker
On the broader subject of vid card reviews at Kyle's, there was a lengthy discussion about his review methodology back in late '99 early 2K here when Kyle was trying to demonstrate the viability of a new feature on a graphics board and totally blew it(trying to demonstrate T&L's ability but he ran a fill limited situation). This in itself wasn't too big of an issue however when he was informed of how to display more accurate results for what he was trying to do(which only required that he lower the resolution) he refused to do so and made no attempt to ever rectify the misconception that he had given a lot of people(IIRC it was myself and LeoV that really went off about it).

oh sure we all have our faults and i have stumbled across Kyle's before on my own; you can go do a search over there for posts with his name in the text and my same user name as here if you want to see for yourself. also, my pseudonym "reefer kills" was banned from their forums for speaking up as well.

regardless, i think Brent has come a long way in his reviewing skills lately and Sean is doing alright himself, so i find any dismissal of the reviews of the [ H ] based on Kyle's history rather far fetched.


lol, you have to trick the forums into useing the [ H ] on Anand, is it a conspericy?

 

[Rogozhin]

yea

what the hell is with the [ H ]

it wouldn't let me post with the real version in.

Kyle has always been an nvidia man

Go watch his footage of the gfx release, they had an auditorium full of screaming nvidiots all stoked about the FX, this was before he did his review.

It was really sad to see that and then read his review of that card.

Such a letdown.

rogo