ATi & nVidia Anti-Aliasing Performance
Anti-Aliasing quality and performance compared while using anisotropic filtering

Dezember 10, 2003 / by Leonidas / page 3 of 10 / translated by 3DCenter Translation Team

   Quality Comparison

What we've outlined so far about the technical specifications of ATi's and nVidia's respective anti-aliasing modes only serves one purpose: to show that these modes cannot be directly compared. The edge quality of the various modes is simply too different, particularly so with 4x anti-aliasing. That's why we won't draw any such direct comparisons when doing our benchmarks, but we'll rather benchmark all anti-aliasing modes we can and then look how it all stacks up.

We'd like to apply a certain quality preference, even though it is, of course, according to our own subjective perception of quality. The following chart presents a relative ordering of both ATi's and nVidia's anti-aliasing modes according to their smoothing quality. 100 per cent would be a theoretical peak that only something on the order of 64x anti-aliasing could hope to achieve. The quality valuation is based on both technical specs (EER and sampling grid) and the comparison screen shots we've already shown (for nVidia we're again exclusively using aTuner's naming scheme):

Modes	ATi	nVidia
1x2	-	25%
2x1	-	25%
2x	40%	40%
4x	70%	50%
4xS	-	60%
4xSS	-	50%
6x	85%	55%
6xS	-	65%
8x	-	60%
8xS	-	80%
12x	-	70%
16x	-	85%

The influence of gamma correction on anti-aliasing, that's exclusive to the ATi chips, has been disregarded for this comparison. It's hard to express the effect in terms of edge smoothing quality, but it's, of course, an additional advantage for ATi's anti-aliasing modes. To a certain extent, nVidia's supersampling hybrid modes can counter this effect, because they automatically enhance texture quality (while multisampling strictly affects edges). But then, this effect wouldn't be right to place in the above chart either. It doesn't have anything to do with edge smoothing quality, and that's what we're looking for (it's rather an appreciated byproduct).

For the performance comparison between ATi's and nVidia's anti-aliasing modes, we ended up picking the following: ATi 2x, 4x and 6x, and nVidia 2x, 4x, 4xS, 8xS and 12x. We also benchmarked nVidia's 8x mode, but as its quality doesn't compare favorably to 4xS while its performance is significantly worse, these numbers didn't make it into the comparison. They are, however, present in the appendix.

   Performance Comparison

We used the same test system as we did here, the only difference is that we've installed DirectX 9.0b. We've used Catalyst 3.6 for ATi cards and Detonator 45.23 for nVidia cards. Both aren't exactly the most recent versions, but as performance increases in later drivers are largely constrained to shader limited benchmarks (which we don't use here), we hope to get away with this.

We used Aquanox 2, Comanche 4, Dungeon Siege, Max Payne, Serious Sam: The Second Encounter, Star Trek: Elite Force II, Unreal Tournament 2003 and Wolfenstein: Enemy Territory for benchmarking. Initially we planned on using No One Lives Forever 2 and Splinter Cell as well but had to skip these, because some of the anti-aliasing modes we wanted to use didn't work with these games. As noted in the introduction, throughout all benchmarks, we've used 16x anisotropic filtering on ATi cards (full trilinear has been forced with rTool) and 8x anisotropic filtering on nVidia cards.

The main contenders were a Radeon 9800 Pro 256MB and a GeforceFX 5900 Ultra 256MB. To compare 128MB versus 256MB we added a Radeon 9800 Pro 128MB and a GeforceFX 5900 128MB. For cross comparison to ATi's and nVidia's respective lesser performance cards we also added a Radeon 9700 128MB and a Geforce FX 5800 128MB. Initially we wanted to have a more comprehensive list of cards, full results of (almost) every high end card from ATi and nVidia are in the appendix.

So we don't want to step through single benchmarks and instead skip directly to the overall performance. For this purpose we simply averaged all benchmark results generated with the aforementioned eight games. Because none of the games in our suite are high fps games (100+), this should be sufficient. For the two Direct3D-only anti-aliasing modes offered by nVidia (4xS and 12x), we only took the Direct3D games for calculating the average (5x Direct3D, 3xOpenGL).