Why neither ATI nor nVidia use Fragment AA?

[Kreshna Aryaguna Nurzaman]

Well, I'm talking about the 16x Fragment Anti-Aliasing originally used by Matrox Parhelia. The AA method only takes "fragment pixels" (pixels on the edge of an object) and then collects 16 sub-pixel samples for AA purpose.

Theoritically, this method means way bigger sample size with less fill rate penalty, especially since fragment pixels typically account for less than ten percent of the total amount of pixels displayed on the scene.

So theoritically, the Fragment AA method allows bigger pixel size with less performance drawback.

Judging from the picture below, this method produces razor sharp edges, which is very clean. I think this is because sample size is always more important than sampling pattern when it goes to AA quality, and edge AA allows way bigger sample size (16x compared to typical 4x or 8x used in conventional FSAA method).


(image copied from First Look: Matrox's Parhelia-512 graphic processor article on Tech Report (Page 9)).

The question is..... why neither ATI nor nVidia uses this AA method? Today's video cards have much more processing power than those of Parhelia's era, so imagine the beauty of 64x Fragment AA or such. Or how about enabling AA with very little performance penalty?

What are the problems with this 16x Fragment AA method, so neither ATI nor nVidia have adopted it? Okay, granted Fragment AA does not eliminate texture shimmering (unlike conventional FSAA, which elminates both edge aliasing and texture crawling), but there is Anisotropic Filtering for such purpose.

I have to admit that I missed Parhelia the first time it came around, but did anyone ever have that card? Based on your experience, did you find many problems with Fragment AA so it's actually not worth it?


EDIT: I also wonder why didn't 3dfx use this method with Voodoo5?

[Obi-Wan Kenobi]

Hello Kreshna!:)

Well I have to tell you some good news, even that this topic is somewhat on the older side, I have a Matrox Parhelia AGP 128MB Rev.A 2502 and I tested FAA x16 in numerous game titles and I must say FAA x16 has the performance of FSAA x4 but the image quality of FSAA x8.

The P512 GPU does this very well with the latest DirectX9 beta drivers from February 2007.

I ran the Reef Demo with 1 Billion Colors + FAA x16 enabled on an AthlonXP 2700+ @ 2167Mhz and it rendered just fine, games like Quake III Arena, Descent 3 1.4 + mercenary and Pyromania 1.5 also worked very well and DXX_Rebirth a Descent 1/2 project I have been beta testing for also works just fine with the Parhelia.
here is my Matrox Parhelia Gallery:
http://www.3dfx.ch/gallery/v/3dfx_collectors/collectors_goldleader/matrox/P512-128/

top side:

High-Res:
http://www.3dfx.ch/gallery/d/35081-1/Matrox+Parhelia+AGP+x4+128MB+256Bit+DDR+Rev_A+2502+Top.JPG

rear side:

High-Res:
http://www.3dfx.ch/gallery/d/35075-1/Matrox+Parhelia+AGP+x4+128MB+256Bit+DDR+Rev_A+2502+Rear.JPG

card + box shot:

High-Res:
http://www.3dfx.ch/gallery/d/35084-1/Angled+P512+like+the+boxart_s+P512+pic.JPG


Here some shots I took from the Matrox Parhelia Reef Demo ver 1.1:


High-Res:
http://www.3dfx.ch/gallery/d/34438-3/reef+demo+P512+FAAx16+a.png


High-Res:
http://www.3dfx.ch/gallery/d/34467-2/reef+demo+P512+FAAx16+c.png


High-Res:
http://www.3dfx.ch/gallery/d/34470-2/reef+demo+P512+FAAx16+d.png


High-Res:
http://www.3dfx.ch/gallery/d/34482-2/reef+demo+P512+FAAx16+e.png

What the P512 GPU also had was Tesselation 1.2, this was an improved way of ATi's Truform which was also a very basic variant of Tesselation.

Tesselation 1.2 with the P512 also shows perfect formed 3D animations like them fish and that shark, they have no more bumpy surfaces, this is because of the active tesselator which is at work there.

And this technique has been refreshed by ATi's HD 4800 series and 5800 series, ATi it's self calls it the Tesselator, the R600 from the HD 2900 series was the first VPU from ATI to have retaken this idea back to the gaming platfrom, ATi is also the only Graphics vendor known in the world that is usinmg the newer way of Tesselation for DX10.0, 10.1 and 11.