Benchmarks - Anisotropic Filtering and FSAA Performance

Here we'll take a look at the performances of the X700's under some Anisotropic Filtering and FSAA modes. For these tests we'll use 3DMark03's Game Test 2, "The Battle for Proxycon", at 1024x768, which is a fairly graphically intense environment.

 

 

X700 XT 54.9 51.3 48.3 45.0 41.8
X700 PRO 47.1 44.1 41.2 38.3 34.9
 
X700 XT -6.6% -12.0% -18.0% -23.9%
X700 PRO -6.4% -12.5% -18.7% -25.9%
 
X700 XT -6.6% -5.8% -6.8% -7.1%
X700 PRO -6.4% -6.6% -7.0% -8.9%

As we'd expect, given that both boards are based on the same chip, the relative performance drops for each mode are roughly the same between the PRO and XT, although obviously the XT has a higher overall performance. In both cases the performance drop from each mode is fairly consistently around the 6% mode, except for the transition from 8x to 16x AF, which is a little higher than the rest, certainly in the case of the PRO - this is somewhat curious since, theoretically, there should be less of an increase with higher AF levels since there will be fewer surfaces that would need the full 16x sampling required.

 

 

X700 XT 55.8 41.1 26.2 16.8
X700 PRO 48.3 32.0 20.3 13.9
 
X700 XT -26.3% -53.0% -69.9%
X700 PRO -33.7% -58.0% -71.2%
 
X700 XT -26.3% -36.3% -35.9%
X700 PRO -33.7% -36.6% -31.5%

Looking at the FSAA performance we can see that both the boards appear to be getting their lowest hit at 2X FSAA, especially in the XT's case. 6X FSAA in this case is incurring quite a performance penalty, and it may be the case that the cache sizes on the X700 are more turned towards 2x and 4x FSAA than 6x, given that these boards only have a 128-bit memory bus.

 

 

X700 XT 41.8 31.6 20.8 14.3
X700 PRO 34.9 24.4 16.2 11.7
 
X700 XT -24.4% -50.2% -65.8%
X700 PRO -30.1% -53.6% -66.5%
 
X700 XT -24.4% -34.2% -31.3%
X700 PRO -30.1% -33.6% -27.8%

With 16x AF enabled the trend in performance hit is similar to the case without AF, however obviously the overall performance is lower. The performance degradation for FSAA isn't quite as great in this case as the Anisotropic Filtering is taxing the texture fill-rate more, rather then the bandwidth and the multi-sample fill-rate, hence the two processes can balance each other out a little.