FSAA


FSAA Performance

We'll use the custom level in SS:SE to take a look at the FSAA performance of the GeForce FX 5900 Ultra. We'll take a look at both DirectX and OpenGL as there are some modes that are not available under OpenGL.

None 119.8 109.8 93.8 70.1 53.6
2X 103.1 95.6 82.5 60.8 46.5
Qunicunx 102.5 94.6 81.7 61.1 46.1
4X 101.7 90.8 73.7 51.0 37.1
4xS 97.1 80.1 57.8 38.7 27.4
6xS 90.9 71.2 49.1 31.5 27.9
8X 82.0 60.4 40.2 23.0 27.8
2X -14% -13% -12% -13% -13%
Qunicunx -14% -14% -13% -13% -14%
4X -15% -17% -21% -27% -31%
4xS -19% -27% -38% -45% -49%
6xS -24% -35% -48% -55% N/A
8X -32% -45% -57% -67% N/A

As ever, the performance difference between 2X and Quincunx FSAA remain the same, as they both use the same number of subsamples and Quincunx's post filter is done within the RAMDAC, so there is virtually no performance loss. All the modes beyond 4X are mixed Super and Multi Sampling modes, and thus they have a higher fill-rate requirement than the pure MultiSampling modes, which is why the performance hit begins to rise steeply, as well as the bandwidth requirement they have. Note that 6xS and 8X modes have the same performance at 1600x1200 as 4xS indicating that its actually dropping down to this mode.

None 149.7 145.2 137.0 108.6 82.5
2X 143.1 139.0 127.6 96.6 72.2
Qunicunx 143.4 142.1 128.7 97.3 72.6
4X 144.5 140.0 123.4 90.1 66.9
8X 92.1 66.6 43.8 28.1 19.3
2X -4% -4% -7% -11% -12%
Qunicunx -4% -2% -6% -10% -12%
4X -3% -4% -10% -17% -19%
8X -38% -54% -68% -74% -77%

For 2X, Quincunx and 4X modes the performance drop are lower under OpenGL than the respective modes in OpenGL, though the reverse is true for 8X. 8X mode is still a mixed Super and Multi Sampling method, yet in this instance NVIDIA have managed to make it operate under OpenGL, unlike the xS modes.