ATi & nVidia in driver "optimizations" race

July 26, 2003 / by Leonidas / page 8 of 9

   ATi Optimizations  (cont.)

Given that ATi´s general optimizations of the anisotropic filter also applies to all other applications, we ran the previously used benchmarks except Unreal Tournament 2003 once again with the 16x anisotropic filter. The anisotropic filter was once regularly adjusted over the control panel, which leads to a bilinear/trilinear anisotropic filter and once over rTool 0.9.9.6d, which results in a pure trilinear anisotropic filter:

ATi 03.4 Pentium 4 Northwood 2.53 GHz  -  Radeon 9700 Pro 128MB 1024x768x32  -  16xAF   "Quality" AF per Control Panel   trilinear AF per rTool 3DMark2001 SE 11701 Pts. 12761 Pts. 3DMark03 (Build 330) 3937 Pts. 4026 Pts. AquaNox 2 69,5 fps 64,2 fps Codecreatures 26,1 fps 26,1 fps Comanche 4 42,3 fps 42,6 fps Counter-Strike 118,5 fps 119,9 fps Dungeon Siege 68,7 fps 68,4 fps Max Payne 65,1 fps 62,4 fps NOLF 2 63,8 fps 63,4 fps SSam: TSE 51,9 fps 50,9 fps Splinter Cell 36,7 fps 36,7 fps 20% 40% 60% 80% 100% 120%

The other benchmarks however partly produced very confusing results: The fact that there´s no impact at all on some benchmarks suggests a possible CPU limitation and therefore negligible. But the definitely higher results produced under the two 3DMarks with the assigned high-order filtering are at the moment absolutely not explainable. Perhaps we return again later to this phenomenon, but for the instant we don´t want to keep ourself busy with the two 3DMarks, since both benchmarks are considered by ATi & nVidia to be "in need" for application-specific optimizations and so became uninteresting as comparison instruments between graphics cards anyway.

Apart from these two exceptions at least in AquaNox 2 and Max Payne a few lower results can be observed as they were expected to be seen with the high order filtering produced by rTool. The difference between the bilinear/trilinear anisotropic filter by control panel and the pure trilinear anisotropic filter by rTool is however clearly smaller than the difference measured with the flybys in Unreal Tournament 2003.

Finally there´s a need to check if these observed performance differences also apply to the newer drivers 03.5 and 03.6. For that purpose we ran some selected Unreal Tournament 2003 flyby benchmarks once again to compare these drivers with the 03.4 driver. But there shouldn´t be any real differences, because ATi´s filter optimization is - as shown before - identical within the driver versions 03.2 to 03.6:

ATi 03.4, 03.5 & 03.6 @ UT2003 Pentium 4 Northwood 2.53 GHz  -  Radeon 9700 Pro 128MB 1024x768x32  -  16xAF   "Quality" AF per Control Panel   trilinear AF per rTool Antalus (03.4) 121,5 fps 92,2 fps Antalus (03.5) 121,7 fps 92,2 fps Antalus (03.6) 121,7 fps 92,2 fps Asbestos (03.4) 167,2 fps 132,0 fps Asbestos (03.5) 167,1 fps 132,0 fps Asbestos (03.6) 167,2 fps 132,0 fps Suntemple (03.4) 127,8 fps 106,3 fps Suntemple (03.5) 127,9 fps 106,3 fps Suntemple (03.6) 127,9 fps 106,3 fps 20% 40% 60% 80% 100% 120%

Thus it can be verified that the ascertained performance differences of the driver 03.4 with the anisotropic filter are transferable 1:1 to the newer drivers 03.5 and 03.6. An additional control of real screenshots with the driver version 03.6 brought no new revelations: Also with this driver version no further differences except the bilinear/trilinear anisotropic filter showed up between the filter produced by ATi control panel and that one produced by rTool (therefore we also did without these screenshots).