The new 8x, 12x and 16x Anti-Aliasing on GeForceFX
April 13, 2003 / by aths & Leonidas, translation by Zeckensack & nggalai / page 5 of 6
OpenGL Performance
Once more we'll throw in some benchmark results from Unreal Tournament 2003's Flyby on CTF-Face3, at 1024x768x32 and maximum detail settings. The generally lower scores are not representative of actual in-game performance where OpenGL and Direct3D perform roughly the same. This disparity is related to the fast-running flyby and may as well be disregarded:
|
Unreal Tournament 2003
MaxQuality 1024x768x32 OpenGL | |
Pentium 4N 2.53 GHz · i845PE · 512 MB DDR-RAM · GFFX 5600 Ultra | |
|
22 fps |
44 fps |
66 fps |
88 fps |
110 fps |
132 fps |
| |
This one's obvious: New modes and old modes bear equal names in the Detonator and aTuner. Neither image quality nor performance figures indicate any differences, including the new 8x and 16x settings (both of which are correctly named).
But there's a peculiar anti-aliasing limitation under OpenGL: Supersampling is only applicable if both screen axes are uniformly scaled. The GeForce 1/2 series for example uses 1x2 supersampling under Direct3D, but 1,5x1,5 supersampling under OpenGL. As a consequence 4xS is not available on Geforce 3/4/FX under OpenGL because it has a 1x2 supersampling component. The same logic applies to other modes available for Direct3D.
In theory a "4.5x" or "9x" setting could have been done, but the 1.5 scaling factor in conjunction with multisampling leads to artifacts, similar to those seen with the 12x mode in Direct3D. That's why under OpenGL the next mode beyond 4x is 8xS. As this is exactly the same thing as the Direct3D mode of the same name, it shares its fantastic quality and of course its extreme performance impact, unfortunately limiting its usefulness quite a lot.
The next step up to 16x anti-aliasing is near useless: Because of the switch to a less efficient subpixel grid the "Effective Edge Resolution" (EER) stays at 4x4. The overall quality gets only marginally better, only edges at angles close to 45 degress might see some improvements. We think this hardly justifies the doubled framebuffer footprint compared to 8xS and conclude that 16x under OpenGL is just as irrational as 12x under Direct3D.
