Vertex Shader

See here for details of this test.




 640x480 800x600 1024x768 1280x1024 1600x1200
9700 PRO 15.2 15.1 14.9 14.6 14.2
9700 13.0 12.9 12.7 12.5 12.2
9500 PRO 13.5 13.4 13.2 12.8 12.4
9500 13.4 13.1 12.9 12.3 11.6
9000 PRO 3.8 3.8 3.7 3.7 3.6
Parhelia 6.1 6.1 6.0 5.8 5.6
4600 6.9 6.9 6.8 6.8 6.7
4200 6.3 6.2 6.1 6.1 6.0
 
% Diff from 9700 PRO 640x480 800x600 1024x768 1280x1024 1600x1200
9700 -14% -15% -15% -14% -14%
9500 PRO -11% -11% -11% -12% -13%
9500 -12% -13% -13% -16% -18%
9000 PRO -75% -75% -75% -75% -75%
Parhelia -60% -60% -60% -60% -61%
4600 -55% -54% -54% -53% -53%
4200 -59% -59% -59% -58% -58%

Being a primarily Geometry based test we can see that increasing the resolution affects the test very little, meaning that it is stressing the Vertex Shaders as it should be. All of the R300 based ATI boards are where we would expect them to be, given that they all have 4 Vertex Shader processors. The Radeon 9700 PRO is a little distanced from the others because of its higher clock rate, however, oddly the 9500 PRO performs slightly better than the 9700 which is unexpected as they both have the same theoretical Vertex performance -- presumably the 9500 PRO is making better use of its smaller memory bus than the 9700 is of its wide 256-bit bus.

The rest of the boards are roughly where they are expected to be given their respective clock speeds and number of Vertex units. The obvious exception to this is Parhelia, which is supposed to have four DX9 class Vertex Shaders, and is yet unable to outperform either of the GeForce4 Ti's.

Pixel Shader 2.0

See here for details of this test.





 640x480 800x600 1024x768 1280x1024 1600x1200
9700 PRO 66.5 53.4 40.9 30.1 23.8
9700 57.8 46.4 35.6 26.2 20.7
9500 PRO 53.0 42.2 32.2 23.6 18.7
9500 39.6 30.6 22.8 16.1 12.2
 
% Diff from 9700 PRO 640x480 800x600 1024x768 1280x1024 1600x1200
9700 -13% -13% -13% -13% -13%
9500 PRO -20% -21% -21% -22% -21%
9500 -40% -43% -44% -47% -49%

The new Pixel Shader test in 3DMark03 is done via PS2.0, hence this test is reduced down to the DX9 cards only. The performance spread of cards is relatively predictable, with each of the boards performing roughly where we would expect them.

Unlike the "Mother Nature" test, this test is showing a little bit more of a bandwidth limitation, as the 9500 PRO has dropped behind the 9700 a little, although not by a great margin. It's interesting that this is showing some minor bandwidth limitations as the point of procedural texturing is that the actual texture detail is generated programmatically via a Pixel Shader program, so you may expect this to be shader limited rather than bandwidth limited on the 9500 PRO. However, as mentioned in the Introduction to 3DMark03 a turbulence or noise map is used to sample from, which will require bandwidth. The Fill-rate graph also shows that this isn't entirely Pixel limited either, as most of the boards have quite a steep rise upwards, rather than coming to a plateau.

The 9500, with half the Pixel Shader pipes, illustrates the how much the 8 pixel shader pipes are doing for the other boards, as it is showing itself to be much more fill-rate limited.