Zum 3DCenter Forum

Detonator 52.14 Test & 52.10 Re-Test

October 8, 2003 / by Leonidas / page 5 of 7

   Game Performance  (cont.)

As already mentioned, there is no known possibility to compare the drivers 45.23 and 52.10 with fully identical image quality. We will still attempt an approximation, because we can show the benefit of the new driver "optimizations" by doing so, at least partially. For this, we need games where we can adjust the setting for the anisotropic filter from within the game. For e.g., this are, among the Direct3D games, Aquanox 2 and Unreal Tournament 2003.

Insofar as you did activate the anisotropic filter and set it to 8x in the corresponding game, as well as selecting the Application setting in the control panel, which enables the application to take control of the filter quality, an exemplary filter should be the result. Sadly, this does not quite work out, as the Detonators 52.10 and 52.14 (other than Detonator 45.23) don't use a real trilinear filter, not even in the Application mode, but rather a faked trilinear filter. But at least, with this method, we don't need to worry about the further "optimizations" of the Control panel modes (CP = per Control Panel, App = per Application):

8xAF (App)
8xAF (CP)
8xAF (App)
8xAF (CP)
AquaNox 2 74.0 fps 58.3 fps 80.5 fps 77.4 fps
Serious Sam: TSE (D3D) 56.0 fps 56.1 fps 58.4 fps 59.2 fps
UT2003 - Pyramid2003 44.3 fps 44.3 fps 42.7 fps 45.3 fps
UT2003 - Antalus-Flyby 128.4 fps 128.4 fps 84.4 fps 127.7 fps

These benchmarks deliver some eye-popping results: First, the clearly improved Aquanox 2 benchmarks between versions 45.23 and 52.10/14 seem not to be the result of driver improvements, regardless if legal ones or illegal ones. This is because already with the Detonator 45.23, you can have the good values of the Detonators 52.10 and 52.14 - exactly then, when the game itself takes control of the anisotropic filter setting.

Apparently, the AquaNox 2 programmers themselves have implemented some (legal) optimizations for the anisotropic filter, which are however only activated in the Application mode. If, on the other hand side, the Control Panel controls the setting, we can see an exemplary anisotropic filter when using the Detonator 45.23, but the performance remains clearly below that of the optimized filter of the game itself. At least the difference of 58.3 fps per Control Panel to 74.0 fps per Application mode using Detonator 45.23 in AquaNox 2 is probably caused by the optimization of the anisotropic filter in the game itself - however, this is just a guess on our behalf.

Therefore, alone the difference using the Application mode from 74.0 fps (Detonator 45.23) to 80.5 fps (Detonator 52.14) could make a statement on how much the speed of the game has been increased. Though also here, using the Detonators 52.10 and 52.14, the faked trilinear filter is in effect, as opposed to the real trilinear filter in the Detonator 45.23. It is possible that the performance increase of 6.5 fps is simply on behalf of the faked trilinear filter, because quite apparently AquaNox 2 reacts quite strongly on the alteration of the filter quality. Thus, the ostensible performance boost of the drivers 52.10 and 52.14 is not by means of driver improvements, for the very most at least.

In Unreal Tournament 2003 you can see from the benchmarks how much faster the filter, set by the Control Panel, is in comparison to the filter in Application mode, using the Detonators 52.10 and 52.14: Whopping 51% in the more theoretical Flyby and 6% under real-world conditions in a timedemo. Of course, both of this is only caused by the afore mentioned "Optimizations" of the Control Panel mode.

Likewise, you can also read out of these numbers that the Detonator 45.23 contains an "optimization" which has remained unseen for Unreal 2003: Because the benchmarks in Application mode as well as in Control Panel mode deliver identical results, the Application mode in the Detonator 45.23 using Unreal Tournament 2003 is not in effect, and the game is "optimized" by force even here.

Using drivers from 44.03 to 44.90, we could already see an application-specific "optimization" for Unreal Tournament 2003 back then, which however could be circumvented manually by using the Application mode and setting the anisotropic filter in the game itself (UT2003.ini, see "LevelOfAnisotropy"). The Detonator 45.23 cannot be tricked anymore by this, Unreal Tournament is now always "optimized", no matter what (MouseOver effect using Javascript, alternatively: click opens both screenshots):

MouseOver effect using Javascript, alternative: click opens both screenshots

Thus, the sole remaining possibility to determine the exact performance advantage of the faked trilinear filter of 52.10 and 52.14 are the benchmarks in the Direct3D mode of Serious Sam: TSE. Because both results using the Detonator 45.23 are nearly identical, we can conclude that SS:TSE has no filter optimization in the game itself, or that nVidia is "optimizing" right in the driver - good preconditions for a comparison with Detonator 52.14.

Compared to that one, SS:TSE only gains 4.3% in the Application mode. In case nVidia didn't implement any other improvements other than the improvement of the shader performance (it looks quite like that), these 4.3% can be attributed to the faked trilinear filter of the Detonators 52.10 and 52.14. The additional "optimizations" of the Control Panel mode increase the performance of SS:TSE for another 1.4%. This really isn't much, considering the quite heavy "optimizations", but possibly this may be due to the CPU limitation, which is around 60 fps in this demo on the computer used here.

However, testing SS:TSE alone is not quite enough for making a general statement about how much performance is gained by the faked trilinear filter of the Application mode and the other "optimizations" of the Control Panel mode in the Detonators 52.10 and 52.14. According to the benchmarks above, and carefully guessed, we would say that the performance gain of the Application mode is around 2-4% and the performance gain of the Control panel mode is around 3-6%.

comment this article in our english forum - registration not required Zurück / Back Weiter / Next

nach oben