Benchmarks - Anisotropic Filtering and FSAA Performance

Here we'll take a look at the performance variants for the various modes of FSAA and Anisotropic Filtering on a scene. For this test we're using 3DMark03's Game Test 2, which utilises a number of different rendering features (texturing, shading, stencil buffers etc.), at 1024x768 as at this resolution it is still very graphics intensive.

When looking at the performance drop for each level of Anisotropic Filtering in this test the pattern is much as you would expect - for each level of AF applied the performance drop is less than the previous level performance drop. This is to be expected because the level of filtering is the "Maximum" level of Anisotropic Filtering hence there will be fewer surfaces that actually require 4x AF in relation to 2X and again fewer still that actually require 8x AF, etc.. The total performance drop for going from 1x AF to 16x is 16% in this case.

Not unexpectedly the performance drops for FSAA are quite high with the 6200, and it looks to be due to the bandwidth available. That being said with 4 times the number of samples being drawn for 4x FSAA we would logically expect to see a quarter of the performance if it was entirely bandwidth limited, but this is not the case here - although the Z compression is disabled for 6200 it appears that the colour compression, at least where FSAA is concerned, is still enabled.

With 16x AF applied we are still getting similar actual performances to those without AF, but the relative performance drop for each level of FSAA is reduced as the board is fairly texture fill-rate limited in the first place, thus alleviating some of the bandwidth issues.