What is the best real-time anti-aliasing ever?

[Megadrive1988]

What I am asking here is, what is the "best" anti-aliasing ever done in real-time? I suppose the definition of "best" here would be a combination of at least two things: number of samples (i.e. 2x2, 4x4, 4x) combined with the type of AA. (i.e. super sample or MSAA)

Honestly, I don't know jack about AA. but I am interested I know that it makes things look so good. that's one of the reasons CG looks so great, its got TONS of AA, that cannot be done in realtime. yet.

obviously we have to seperate real-time from prerendered. in CG, there might be dozens of AA samples (i.e. 64x AA ) right? which is not done on a real-time basis of course. then there is the AA performed in real-time. in PC GPUs, older T&L-less 3D accelerators, workstations, flight simulators, highend visualization systems. (SGI, E&S, 3DLabs, Intergraph, etc)

I've never heard of more than 8x8 (or maybe just 4x4) or 16x AA done in real-time. or is there?

part of achieving the "CG look" which is the topic of another thread in the console forum, is getting lots of AA done. right now, the most any consumer GPU does is 8x (NV3X) from what I understand.

The upcoming R420 and NV40 should be a step in the right direction, hopefully. But I am not expecting things to improve enough until we reach the R500 and NV50 generation

 

[bloodbob]

I'm sure you could aa something at like 800x600 SSAA at 1x1 resolution

 

[akira888]

While the Nvidia 2xSSAA by 4xMSAA mode (8X) is the AA setting with the most subsamples/pixel it is far from the best. The reason for this is that the samples are arranged in an ordered grid, which on horizontal and vertical lines can produce ugly gradient step artifacts. The reigning Gold Standard in consumer level AA is Ati's 6x SG-MSAA with automatic gamma correction. Since all of the sample points are on different horizontal and vertical lines on a 12x12 subsample grid, nearly straight triangle edges show less "stair-stepping" than do the comparable Nvidia modes.

 

[Pete]

I'd imagine 3dfx's Voodoo 5 series had the best FSAA (RGSS + temporal AA), and ATi's current Radeon series has the best edge AA (JGMS + gamma correction). I think.

 

[Ailuros]

(IMHO)

Best Supersampling algorithm: 3dfx VSA-100

Best Multisampling algorithm: ATI R3xx

Most innovative algorithm (despite the negatives): Matrox Parhelia

 

[pcchen]

If the 3DLabs Wildcat's "SuperScene" really works as advertised, it could be the currently best real-time FSAA.

 

[GameCat]

Yeah, 3DLabs WildCat wins hands down. First of all the AA is adaptive IIRC so up to 16 samples per pixel depending on how much it needs. Besides that, the samples are jittered (on an ordered grid but offset by a small amount in a random direction) whichis very good at reducing aliasing. I've seen animations f it in action and the AA quality is very good.

 

[Ozymandis]

I might be mis-remembering, but I think my Voodoo5's 4x FSAA was nicer looking overall than 9800pro's 6x. *I think* I can't imagine how good the unreleased model's 8x looked

Of course, the fact that I can run my ATI's AA at great framerates and at good resolutions does tip the scales a bit. The 5500's FSAA just wasn't that fast at 4x

 

[Gnep]

Matrox's FAA 16x would have done it for me in the consumer space.

If it didn't have occasional artifact issues (although apparently most of these are fixed in the Px50 series - although they're not gaming cards ).

Gnep

 

[rashly]

(IMHO)

Best Supersampling algorithm: 3dfx VSA-100

Best Multisampling algorithm: ATI R3xx

Most innovative algorithm (despite the negatives): Matrox Parhelia

i wouldnt call the 3dfx algorithm "best" because of the huge performance drop. i would say that the quality is of the best (especially 6000 8x).

ati's 6x has gamma correction but id give the quality crown to the 6000 based on personal experiences with both cards.

of course id rather have my 9800 doing 6x fsaa than a 6000 doing 8x fsaa. its amazing that something made 3 years ago has the same quality as the newest part today.

 

[Megadrive1988]

Don't forget that non-consumer AA is included too. as long as it's
realtime /interactive, such as highend simulators and visual computing.

certainly the best realtime AA is NOT available in a consumer chip(3Dfx,ATI)

 

[Ailuros]

(IMHO)

Best Supersampling algorithm: 3dfx VSA-100

Best Multisampling algorithm: ATI R3xx

Most innovative algorithm (despite the negatives): Matrox Parhelia

i wouldnt call the 3dfx algorithm "best" because of the huge performance drop. i would say that the quality is of the best (especially 6000 8x).

ati's 6x has gamma correction but id give the quality crown to the 6000 based on personal experiences with both cards.

of course id rather have my 9800 doing 6x fsaa than a 6000 doing 8x fsaa. its amazing that something made 3 years ago has the same quality as the newest part today.

I concentrated mostly on quality and on mainstream consumer purchasable boards. The V5 6k doesn't belong into that frame.

I used the distinction between Multi-/ Supersampling and FAA on purpose. If I would have the choice between playable 8x RGSS in 1024*768*32 and 6x RGMS/16xAF in =/>1280*960*32, I'd of course pick the latter, but that's a matter of personal preference.

 

[WaltC]

i wouldnt call the 3dfx algorithm "best" because of the huge performance drop. i would say that the quality is of the best (especially 6000 8x).

ati's 6x has gamma correction but id give the quality crown to the 6000 based on personal experiences with both cards.

of course id rather have my 9800 doing 6x fsaa than a 6000 doing 8x fsaa. its amazing that something made 3 years ago has the same quality as the newest part today.

The V56K never made it to market, and all that exist today of it are prototypes. So, I think it's a bit unfair to compare it with a production product like R3x0. As others have noted, the V5 5.5K was great at low-res FSAA at the time--and I'd have to say that it wasn't until my R300 last year about this time that I started using FSAA again--post V5. I never used it with the GF3/4 because it just wasn't good enough at the time, by a country mile.

It was kind of interesting, though, to note that in old games like MTM2, it wasn't until ATi supported 16-bit FSAA in the Cats last year and I could view the game (640x480 set resolution) in the 6x FSAA mode, that I saw similar quality to what I remembered with the V5 and that game with FSAA. The V5 FSAA did a great job on removing pixel "shimmer" in games at the time--and I had to go to 6x FSAA on the R300 to get back to that place with that particular game... But at least I was able to do so...unlike with the nVidia products I owned post V5 and pre-R300...

Also, I don't think 3dfx's algorithms were as responsible for FSAA performance as were the physical bandwidth limitations mandated in the V5 design by economics three years ago. Looking at what they had to work with then as compared to now in the way of hardware, I rather think it was 3dfx's algorithms which made the V5 FSAA performance as good as it was with the quality it produced.

I also prefer the R3x0, of course...but have to really hand it to 3dfx for what it did for FSAA in general in getting it kickstarted in the industry.

 

[Ante P]

Don't forget that non-consumer AA is included too. as long as it's
realtime /interactive, such as highend simulators and visual computing.

certainly the best realtime AA is NOT available in a consumer chip(3Dfx,ATI)

es has the ATi r3x0 based RenderBeast which supports 24x AA per card (as in you can scale higher if you want, two cards = 48x AA), should look "pretty" good
theoretically 256 R3x0 boards with 6x AA each = 1536x FSAA should look "ok" me thinks?

V5 6000 8x FSAA looked mighty nice though without some sort of aggressive texture filtering I don't think it would look very good compared to todays standards

other than that I'd say 3dlabs superscene 16x

personally I think I'm quite alright for the time being with ATis 6x FSAA at 1600x1200, I don't need more than that
some better texture filtering would be nice though

 

[Simon F]

I would think that the SGI machines that implemented a version of Schilling's EXACT AA algorithm would have (usually) produced a pretty good result.

 

[Arun]

You haven't seen anything yet in that department
( I'm hinting at the NV40 here )

I'm not sure I can leak that thing, so I will if I can - otherwise, well, maybe the source will leak it himself, or it won't get leaked ( the reliability of the information is, I would say, Kinda High - if you're familiar with GPU:RW reliability rating system )


Uttar

 

[jvd]

You haven't seen anything yet in that department
( I'm hinting at the NV40 here )

I'm not sure I can leak that thing, so I will if I can - otherwise, well, maybe the source will leak it himself, or it won't get leaked ( the reliability of the information is, I would say, Kinda High - if you're familiar with GPU:RW reliability rating system )


Uttar

Sorry man but I can't get hyped up for a new nvidia product. Its like drinking sour milk. Before you drink milk again u will allways smell the milk to make sure its okay. With another nvidia card I need to wait months to make sure there is no cheating and the features that are suposed to be there are there.

 

[MaxSt]

Some time ago I decided to make 16xFSAA in software - just to see how good it looks. It was kinda fun...

http://www.hiend3d.com/16xFSAA.html

MaxSt.

 

[demalion]

You haven't seen anything yet in that department
( I'm hinting at the NV40 here )
...

Yes, I noticed your description, but the way you described the randomness doesn't look encouraging. What is the grid resolution of the ordered grid being used? Where is gamma correction in the picture, absent or "close to the vest" and just refrained from the leak? How many samples, at what performance cost, and with what sample position selection criteria and how is the "randomness" applied?

Most likely best case seems like it will resemble Superscene, but without the answers to the above that doesn't tell us much yet. Though, arguably, it confirms the most important thing of all: better visual quality than NV3x.