FSAA

FSAA Performance

For the purposes of testing FSAA performance we'll once again use Serious Sam: Second Encounter's Citadel demo. The 'Extreme' Addon was used but with Anisotropic filtering disabled. First we'll look at all the available modes under DirectX.

Testing the various AA modes available we can see that 2X FSAA and Quincunx offer a very small performance drop, with at worst only 9% down from non-AA rendering. NV30 features 4:1 colour compression which should allow the FSAA modes to run more optimally as they won’t require as much bandwidth as normal FSAA rendering would.

As we move up to 4X mode we can see that the performance penalty is getting a little greater. The maximum hit here is 51%, and considering that for 4X FSAA without compression would be 75% we can see that it's having some effect, though it doesn’t appear to be too close to its 4:1 maximum compression ratio.

Moving up from 4X the GeForce FX moves into the mixed Super/Multisampling modes the performance penalty become a little greater. With 4xS the performance hit is a little greater than 4X because, although there are the same number of pixel samples, two pixel pipelines are being used to provide two sub samples of each pixel, meaning that the fill-rate is cut in half. Likewise, with 6xS, the fill-rate is cut down to a third of the non-AA rate, as now three pipelines are used to generate each the six subsamples. 8xS only cuts the available fill-rate in half, as it uses two pixel pipes, each producing 4X MSAA, though the extra bandwidth required for all 8 samples means there is a further performance degradation.

The numbers above are from the modes running under the OpenGL renderer of SS:SE. The performance differences are in line with those of the DirectX modes -- they are a little higher, but then the performance of this title under OpenGL is better in the first place.