GFFX Filtering Performance Options

[Brent]

regarding Aggressive, Balanced and Application settings

I'm really getting into testing this now, in fact i may write up an article about it. I think this deserves some attention. I know Dave has taken a look at this in his GFFX Preview as well as the preview at 3DVelocity already.

Doing some prelim testing today I can definitely see that with No AF enabled and Trilinear enabled in the games Aggressive mode looks a lot like Bilinear filtering. Balanced mode is better then Aggressive but still not perfect, you can see the boundries between mip-maps. However Application does look very good with very smooth transitions between mipmaps, in Serious Sam 2.

In UT2k3 its not as noticeable as in Serious Sam 2 but looking closely you can see the same results.

I'm going to fire up quake3 and look at that in a bit as well.

I also just did some performance numbers in UT2K3

I ran our benchmarker in 1280x1024 and 1600x1200 in NoAA/NoAF comparing Aggressive, Balanced, and Application on the GFFX at 500/500

I saw big performance differences between each with Application being teh slowest and Aggressive the fastest

Example

Antalus
1280x1024 Aggressive - 148.3 FPS
1280x1024 Balanced - 134.2 FPS
1280x1024 Application - 92.8 FPS

1600x1200 Aggressive - 108.5 FPS
1600x1200 Balanced - 95.5 FPS
1600x1200 Application - 65 FPS


Now how this compares to default R300/350 Trilinear I haven't looked like yet but am getting there.

Also how this compares with AF enabled I haven't looked at yet either.

Anyways none of this is conclusive yet on my part so PLEASE DON'T QUOTE ME ON ANYTHING!

Thanks

Just thought I'd post my own personaly findings so far on this.

The important part is most are using the default setting for benchmark comparisons, and Balanced is the default setting which looks like it may not match up apples to apples with the r300/350's default trilinear with no af.

 

[Natoma]

While the situation has improved over the past year or so, I still get the strong impression from the "mega websites" such as Anandtech, Tomshardware, and HardOCP that apples-to-apples image comparisons aren't taken as seriously (though to be fair brent, I liked the fact that you discarded the GFFX 2x scores in your Radeon9800 review, due to the fact that 2x does nothing on that card, for instance). Imo, today's graphics cards should *only* be benchmarked with AA+AF.

Why should anyone in this day and age spend $300-$500 on a videocard to run without AA+AF? We've finally gotten past the need to benchmark at 800x600 and 640x480 because the cards today are fast enough. Frankly I believe they're fast enough now to benchmark at 1024x768 with AA and AF enabled as well.

I've gone back over the Anandtech review and noticed that unfortunately they only benchmarked using "Balanced". No use of "Application." I hope that this changes in the future, for all reviews, because I believe it is utterly important to hold not only FPS, but IQ, up as the standard bearer for videocard quality.

It's unfortunate imo that something like this project that you're undertaking Brent is being done not as part of the initial review, but in the aftermath. I feel that this should have been part of all of the initial reviews of this card. I know that the need and the urgency to get a review out "in time" sometimes causes people to gloss over certain aspects. But in most cases, the first review is how people form their lasting opinions of a product. Thus the first review is *utterly* important, and should be treated as such, and not as a "Well we have to get this out at the same time as our competitors!!" type of event.

I think this by itself is a mere reflection of how far the quality of reviews needs to go to truly be useful.

 

[tEd]

The important part is most are using the default setting for benchmark comparisons, and Balanced is the default setting which looks like it may not match up apples to apples with the r300/350's default trilinear with no af.

...talking about re-run all the no AF benches out there...gffx could look very bad if you compare apples to apples

 

[Joe DeFuria]

To be fair, NONE of the aniso modes are apples to apples. nVidia's FX "application" mode is better quality than ATI's quality trilinear.

They're both doing trilinear, but ATI has a lower quality aniso algorithm than nVidia's "application" setting. Generally speaking, quality between all modes goes something like this (From best to worst):

FX Application --> ATI Quality --> FX Balanced --> ATI Performance --> FX aggressive.

FX Aggressive is really in a quality league of it's own...in a bad way. It really isn't fair to compare that to anything.

But ATI's Quality mode falls somewhere between FX's Application, and FX's balanced...

 

[JF_Aidan_Pryde]

One way would be to bench both cards under conditions X, Y and Z such that they both get about C frames a second. Where C is the desired frames per second for that application (30FPS for RPG, 60FPS for racing/sim, 90FPS for First person shooters just as suggestions). And X, Y, Z are Resolution, ansio and AA respectively. This way, since both are getting roughly C frames a second, all you have to do is compare the image quality of a given frame to determine the better one.

 

[pelly]

As Brent mentioned above, we are taking an in-depth look at this issue. You can expect to see this covered in to a certain level within the "full" review of the GeForce FX 5600 Ultra and 5200 Ultra on Monday...However, Brent will be bringing the real depth to this subject with a dedicated article ( if he can find the time...we keep that guy busy! )

8)

 

[jb]

Hi ya Pelly welcome to the gang over here.

 

[tEd]

To be fair, NONE of the aniso modes are apples to apples. nVidia's FX "application" mode is better quality than ATI's quality trilinear.

They're both doing trilinear, but ATI has a lower quality aniso algorithm than nVidia's "application" setting. Generally speaking, quality between all modes goes something like this (From best to worst):

FX Application --> ATI Quality --> FX Balanced --> ATI Performance --> FX aggressive.

FX Aggressive is really in a quality league of it's own...in a bad way. It really isn't fair to compare that to anything.

But ATI's Quality mode falls somewhere between FX's Application, and FX's balanced...

sure not talking about AF.

Just to give a impression here.

Anands ultra preview ->~50% benches without AF
toms ultra preview ->~71% benches without AF

it's about people getting a wrong picture.

 

[Randell]

Its not just about comparison with R300/350. Its about comparison with prior generation hardware as well. If the reviewer isnt careful and benches Gf4 with default true trilinear settings compared to FX Balanced, well as I said elsewhere, fiddling with default LOD settings is sooo 1999..

 

[LeStoffer]

Balanced mode is better then Aggressive but still not perfect, you can see the boundries between mip-maps. However Application does look very good with very smooth transitions between mipmaps, in Serious Sam 2.

Balanced mode is key like Quality (tri) AF on the R3x0. Brent, could you inspect in motion whether the boundries between mip-maps are easy/hard to notice or annoying/okay?

BTW: I'm very please that both nVidia and ATI has made this step to give us good and still fast AF. I'm actually a bit surprised here since AF is something the average end user don't care much about (and I would guess developers didn't bug them much either).

 

[Sharkfood]

I would just settle for a website that runs benchmarks with at least the same level of standards.. and provide IQ shots in a number of different titles and angles of what was benchmarked. This takes the burden off the reviewer to come up with some subjective baseline while showing the readers how close or how far apart the test samples are from equivalency.

As far as same level of standards- if a videocard has to be run over and over and over again in order to get the desired performance, yet the other product is just run once... these results should be thrown out. In the case of the NV30 review samples, they are providing mythical performance results as several sources have cited these boards downclock after several minutes from temperature and do so in the middle of benchmarks often. If this level of speciality is going to be used for one product, then the reviewer should also take the same committment to overclock the competing product in such a way that it's average maintainable clock speed has similar time restraints.

It's absolutely foolish to post performance results with one product that cannot maintain the clockspeed posted while another runs at it's given clockspeed 24/7 with no issues. It's equally foolish when one product's default settings benchmarked look like

 

[Xmas]

It's equally foolish when one product's default settings benchmarked look like

I'm sure you can find some visual glitches on every card. Not like it looks like that everywhere.

 

[madshi]

I'm really getting into testing this now, in fact i may write up an article about it. I think this deserves some attention.

Thanks, Brent. Looking forward to that article.

While the situation has improved over the past year or so, I still get the strong impression from the "mega websites" such as Anandtech, Tomshardware, and HardOCP that apples-to-apples image comparisons aren't taken as seriously

I must defend Anandtech this time. Well, they didn't really understand the differences between all the AF modes (trilinear yes/no etc), but at least they tried to compare the image quality and ended up with benching GeForceFX balanced against ATI performance *and* ATI quality. That's more or less fair in my opinion, because GeForceFX balanced is in terms of image quality somewhere between ATI performance and ATI quality.

Personally, I'd bench application 8x AF on the GeForceFX against quality 8x AF on the R3x0, though. I think those are the 2 modes which can be compared best. Yes, the GeForceFX has a bit better image quality here, but only on certain angles. If you can't bench it this way because of the angle problem, then you can't bench 2x FSAA, either, because the R3x0 has AA gamma correction, while the GeForceFX has not. So if everyone is comparing 2x FSAA benches (but saying in the text that the R3x0 2x AA looks slightly better) they should also compare 8x AF in the best possible mode (and say in the text that the R3x0 AF is inferior at certain angles). Well, at least that's how I see it.

 

[Brent]

Balanced mode is better then Aggressive but still not perfect, you can see the boundries between mip-maps. However Application does look very good with very smooth transitions between mipmaps, in Serious Sam 2.

Balanced mode is key like Quality (tri) AF on the R3x0. Brent, could you inspect in motion whether the boundries between mip-maps are easy/hard to notice or annoying/okay?

Yep, i have looked at it in motion and in motion is where you can notice the differences the most rather then just a static screenshot

clearly in SS2 Application is superior to Balanced and matches ATI's default Tri with no af