Why AA on Geforce FX doesn't look as good as ATI ?

[gryfon]

Hello guys, i'm new to this "FSAA" thing (just jumped from geforce 2 mx to geforce fx 5900), and i'm confused, because in all review site, the screenshot always show that radeon has somewhat better AA and less performance hit while geforce AA isn't as good as ATI and takes more performance hit.
After reading mixed things, i knew a bit, some site says that's because ATI use rotated grid while NVIDIA use ordered. my question is :

1. Why does NVIDIA use ordered instead of rotated ? Is it the hardware, so they cannot change it by driver ? if not because the hardware, can i expect future nvidia driver offer better 4xAA method than they have now, to become at least as good as radeon does (in terms of quality and speed) ?

2. Why the performance hit using ordered grid in NVIDIA card is bigger than ATI ? Is it because they used ordered grid or b'cause ATI's hardware or driver is better ?

3. Why i always see 4x and 2x FSAA used in comparison? 4xs 6x 8x is not used, why ? is it because they're not used widely ? or just takes much more performance hit ?

4. Also in HardOCP i saw an article that shows 4xs is good on nvidia card, but most site that reviews geforceFX or radeon 9800/9600 only shows comparison for 2x or 4xAA, why ?

sorry for the long post, i really need to know

Thanx for any explanation

 

[Matthias99]

1) That's how they designed the cards. It's a hardware limitation of the GeForce cards; it's not like ATI can switch between them either. I'm sure they could do it in software, but that would never be fast enough for practical use. We'll have to see what they do with NV40.

2) Best anyone can tell is that ATI's architecture is just better designed in terms of doing FSAA. Since neither company is terribly forthcoming about publishing all the technical details about how the architecture works internally, it's hard to say definitively where the problem lies.

3) Well, the biggest reason (I would imagine) is that both NVIDIA and ATI can do 4X, and then they diverge. The GeForce cards go to 4xs, and the ATI cards do 6X as their next (and best) AA settings. No cards do 8X that I'm aware of. So it wouldn't be a direct comparison if you did 6X versus 4xs.

4) See answer to #3. Also, the performance hit for 4xs compared to 4x on NVIDIA hardware is pretty big, IIRC.

 

[Genx87]

You should read the IQ article on anandtechs front page. It is pretty interesting to say the least.

 

[Pete]

Well, this stupid browser lost my post, and I don't feel like retyping my whole answer, so I'll summarize:

1. It's hardware. ATi just spent more time on AA quality as well as speed in this gen. The FX's AA quality looks very similar to the GF4's. There may be a few differences (witness the trouble getting AA'ed screenshots in early 5800 reviews), but the quality seems the same.

2. I think it's just hardware, meaning nV can probably spend more transistors on AA in their next part and get up to speed.

3. Probably b/c reviewers are time-constrained (or just lazy). Brent at [ H ] was one of the few who did more in-depth AA analysis with comparative AA screens in UT2K3 and other games in his early 9700/5800 reviews, IIRC. 3DVelocity also had an early 9700 or 5800 review that had extensive AF & AA screens.

4. Probably b/c using 4xS leads to a huge framerate hit, and most reviews are, sadly, still centered around who scores highest, rather than who looks better at, say, ~60fps.

If you want a great explanation and clear illustrations of JG (jittered-grid, of which rotated-grid is a type) vs. OG AA, Google for SSAA-Analyzed.pdf. This is an abbreviated version--3dfx used the illustrations of the previous whitepaper in this presentation of the V5's AA, if you want the differences in easier-to-digest form. Good stuff. B3D also offers a nice illustration of the benefit of gamma correction.

Derek's coverage of AA in his IQ analysis article was very brief (and disappointingly incomplete, but I'll leave that for another thread :\). You'll find a clearer and more in-depth explanation in the 3dfx pdf's I linked above.

 

[gryfon]

thanx for explanation guys, things been lighted up a bit now, though i'm a bit dissapointed that fsaa can't be as good as ati imo, but the nvidia cards still doing great. And yeah genx, i read anand article about IQ, it does interesting.

 

[McArra]

Anand's IQ review was quite dissapointing for me. I thought it would be longer and more in depth.

 

[Pete]

Did I miss a thread on it, or is everyone too busy to comment or just keeping quiet? (In fact, I wrote up a whole pseudo-rant on its shortcomings, but I'm not sure if it's worth fanning any flames by posting such a hastily-written response. I may just let it slide off people's radar.)

 

[gryfon]

may i know why you said that the review is dissapointing McArra ?

 

[BenSkywalker]

A few things real quick, nVidia uses rotated grid for their 2x AA. 4xS combines multi sampling and super sampling which has an advantage over anything ATi has in older titles in particular. Super sampling helps reduce/eliminate texture aliasing and lots of older titles benefit from it, ATi doesn't have any super sampling options for their current boards as of now. 8X is an on again off again option for nV parts, it looks like as of now they won't be enabling it again for some time. It may work nicely for games that are quite old, but the performance hit is too large for them to keep enabled in terms of PR(it is significantly slower then ATi's 6x mode). 8x and 6x AA are largely not that beneficial, if you can run 1600x1200 w/4x AA even on nV parts noticeable aliasing is reduced to the point of being very hard to detect. They are superior modes, and in lower resolutions you can notice it, but you are pretty much always better off using a higher resolution with lower AA settings(exception being LCDs where many have a native res of 1024x768 and it would be of benefit).

 

[BFG10K]

Don't forget that ATi's AA has the advantage of smoothing horizontal and vertical lines much better than nVidia's method, and it also has the ability for its sampling pattern to be reprogrammed by developers if they need something different for their game (eg Valve is using this so it'll correctly work on their packed textures).

 

[Shamrock]

you want in depth? you got it...go read this over at NVNews.net

http://www.nvnews.net/vbulletin/showthread.php?s=&threadid=22166

 

[Rogodin2]

I started that thread over at nvnews because the program that MikeC linked to sparked my interest-it's been followed up by some very valid observations and the other tool that mikec hase found.

In gaming situations I've not been impressed with the fx59xx line (or any other nvidia card)s' implementation of AA-especially in fligh sims where AC SIL SA is VERY important.

Just an observation.

rogo

 

[Shamrock]

i know who started it, but it is still in depth, none the less

 

[McArra]

Originally posted by: gryfon
may i know why you said that the review is dissapointing McArra ?

I was especting more. They told they were doing it a long time ago and I was expecting more pages with more comparisons and screenshots and even some performance comparisons depending on IQ, not on settings. I think I've read better comparisons but thtas just me.

 

[TimisoaraKill]

Nvidia AA has the same quality like ATI AA , belive me , i know a guy who work for Hercules and he still hawe a lot of friends who work for Nvidia , they where friends when Hercules worked with NVDA and they still friends now.
The guy say that Nvidia reduced in fact the 4 AA to be 2 AA and so on to gain more FPS , so when you see on your desk that you run 4 AA ,.... well in fact you run 2 AA , is only a strategy go gain more speed , a dumb one but this does not bother the average costumer the Nvidia employee says .

 

[gryfon]

Nvidia AA has the same quality like ATI AA , belive me ,

i'd like to, but in 4xAA 1024 x 768 i can see the difference quite clearly, or is it just me ?

i know a guy who work for Hercules and he still hawe a lot of friends who work for Nvidia , they where friends when Hercules worked with NVDA and they still friends now.
The guy say that Nvidia reduced in fact the 4 AA to be 2 AA and so on to gain more FPS , so when you see on your desk that you run 4 AA ,.... well in fact you run 2 AA , is only a strategy go gain more speed , a dumb one but this does not bother the average costumer the Nvidia employee says .

So you're saying that on nvidia, 4xAA = 2XAA ? correct me if i'm wrong, but i still see noticeable performace drop going from 2xAA to 4xAA?

According to BenSkyWalker, the 2xAA mode on nvidia is rotated grid (i also remember reading that somewhere), but why the 4xAA is ordered not rotated ? Is it b'cause the hardware needed for 2xAA RG with 4xAA RG is completely different ? I mean, if it somewhat capable of rotated on 2xAA, why not on 4x ?

And also, can you guys tell me performance drop if i enable 2xAA both on nvidia & ATI card compared to disabled AA ? B'cause in 2xAA both are doing rotated grid, so the comparison must be somewhat equal, aren't they ? most sites that review both card only shows non AA and 4xAA + 8xAF, but not 2xAA

thanx guys.

 

[Pete]

This is what Derek's IQ article's AA section should have resembled. The article's in German, but the pictures are universal (it should be translated to English soon, though). AT's IQ article was too long in coming for what we got, IMO. I can't believe we waited a month or so only to read an article that repeated an IHV's assertions without follow-up interviews or analysis (AM3, Halo).

The NVN thread is interesting, but there's some incorrect info in there that's not been corrected. For instance, Edge seems to be under the impression that nV AA's alpha textures with its regular modes, which is false. nV only AA's alpha textures with its mixed AA modes, which combine SS with the now-standard MS AA to AA textures (alpha and every other kind) as well as polygon intersections. nV's mixed modes, OTOH, also incur a large performance hit. All credit to nV in that its AA modes are more considerate of past games, but ATi's AA is more forward-looking--better and faster at the same settings. Also, I believe you can use alpha blends in place of alpha textures to avoid aliasing altogether with partially-transparent textures, so the issue of alpha textures would seem to be less important with newer games.

gstanford also forgets that nV wasn't able to correctly display AA in screenshots of the 5800, and thus required a special version of HyperSnap that basically applied AA outside of the game. So while what you see may be what you get, it's not direct.

The difference between PrntScrn and FRAPS caps that Edge exposed is interesting, though.

gryfon, both do jittered grid (it may be rotated, but I'm not sure) in 2x mode, but ATi still has gamma correction. Nevertheless, 3DC thinks ATi's and nV's 2x modes are equal, interestingly enough.

 

[gryfon]

gryfon, both do jittered grid (it may be rotated, but I'm not sure) in 2x mode, but ATi still has gamma correction. Nevertheless, 3DC thinks ATi's and nV's 2x modes are equal, interestingly enough.

I assume you can read german Pete? so 3DC thinks ATI's and nV's 2x modes are equal ? in terms of quality or performance or both ? translator makes me confuse

in 3DC http://www.3dcenter.de/artikel/ati_nvidia_aa_performance/index2.php, the 2x AA image seems to be rotated MS (the two green dots is not aligned both in vertical and horizontal), am i right ?

 

[Gstanfor]

Originally posted by: Pete
This is what Derek's IQ article's AA section should have resembled. The article's in German, but the pictures are universal (it should be translated to English soon, though). AT's IQ article was too long in coming for what we got, IMO. I can't believe we waited a month or so only to read an article that repeated an IHV's assertions without follow-up interviews or analysis (AM3, Halo).

The NVN thread is interesting, but there's some incorrect info in there that's not been corrected. For instance, Edge seems to be under the impression that nV AA's alpha textures with its regular modes, which is false. nV only AA's alpha textures with its mixed AA modes, which combine SS with the now-standard MS AA to AA textures (alpha and every other kind) as well as polygon intersections. nV's mixed modes, OTOH, also incur a large performance hit. All credit to nV in that its AA modes are more considerate of past games, but ATi's AA is more forward-looking--better and faster at the same settings. Also, I believe you can use alpha blends in place of alpha textures to avoid aliasing altogether with partially-transparent textures, so the issue of alpha textures would seem to be less important with newer games.

gstanford also forgets that nV wasn't able to correctly display AA in screenshots of the 5800, and thus required a special version of HyperSnap that basically applied AA outside of the game. So while what you see may be what you get, it's not direct.

The difference between PrntScrn and FRAPS caps that Edge exposed is interesting, though.

gryfon, both do jittered grid (it may be rotated, but I'm not sure) in 2x mode, but ATi still has gamma correction. Nevertheless, 3DC thinks ATi's and nV's 2x modes are equal, interestingly enough.

I can assure that I have not forgotten the troubles the 5800 had initially. However those troubles centred around AA'd images appearing not AA'd or having less AA than they should have had. In the original pictures edge posted, you can clearly see that AA which did not exist on screen or in FRAPS was magically added. Huge difference.

 

[Pete]

gryfon, no, unfortunately I can't read German, but I think the table near the end of the article offers percentages meaning how effective 3DC considers various AA modes. Both nV and ATi's 2x mode gets 40%. There's a much bigger difference in 4x. Yes, the rotated green dots indicate jittered/rotated grid.

Gstanfor, I see. I mainly skimmed through that NVN thread, but I thought you were disparaging post-frame-buffer AA for the ATi cards, which nV did as well. It's interesting that Edge later discovered that his 4200 exhibits the same behavior as ATi. I'm going to have to reread the thread, as the difference between PrntScrn and FRAPS caps is new to me.

BTW, does nV still require (use) special screenshot-capping filters for its AA modes, or did they change the way they perform AA to allow for correct screenshots without post-rendering filtering?

 

[gibhunter]

According to another 5900SE review I read, the new 53 drivers improve AA quality to be almost indistinguishable from ATI's AA quality.

 

[Acanthus]

Play at 2048x1536 and screw AA imho

 

[THUGSROOK]

if you can tell the difference with AA on ~ stop site seeing and play the game!

 

[Gstanfor]

Originally posted by: Pete
gryfon, no, unfortunately I can't read German, but I think the table near the end of the article offers percentages meaning how effective 3DC considers various AA modes. Both nV and ATi's 2x mode gets 40%. There's a much bigger difference in 4x. Yes, the rotated green dots indicate jittered/rotated grid.

Gstanfor, I see. I mainly skimmed through that NVN thread, but I thought you were disparaging post-frame-buffer AA for the ATi cards, which nV did as well. It's interesting that Edge later discovered that his 4200 exhibits the same behavior as ATi. I'm going to have to reread the thread, as the difference between PrntScrn and FRAPS caps is new to me.

BTW, does nV still require (use) special screenshot-capping filters for its AA modes, or did they change the way they perform AA to allow for correct screenshots without post-rendering filtering?

Yes, fraps is still required to correctly capute some NV AA modes. However, when capturing such modes with printscreen, nVidia screenshots will look as if less AA than required was applied, or no AA was applied when compared with the actual image on screen.

With the sample image edge provided, the driver is clearly adding AA that does not exist onscreen (or in a fraps capture) to the image it outputs. The GF4 shows both FRAPS and printscreen not applying AA. ATi's drivers are the only ones magically adding in non existant AA in that example.

 

[Pete]

Originally posted by: gibhunter
According to another 5900SE review I read, the new 53 drivers improve AA quality to be almost indistinguishable from ATI's AA quality.

I'm not sure how that's possible--they'd have to magically add gamma correction and (perhaps not-so-magically) rotate their 4x mode. Link, please?