As we saw with the earlier 3DMark2001SE theoretical tests the difference in performance between the 6800 Ultra and 5950 is lower than the pure clock rate differences would suggest, which again may indicate that the NV35 pipeline still had some dedicated fixed function geometry processing whereas the NV40 pipeline doesn't. The performance difference between the 6800 Ultra and 5950 in both the VS1.1 and VS2.0 tests are inline with the theoretical performance differences. We see, however, that the 6800 Ultra loses less performance than the 5950, relatively, when branching is used.
Rightmark has Vertex Shader 3.0 tests, but as we neither have DirectX9.0c nor NVIDIA drivers that expose NV40's shader 3.0 capabilities yet, we are not able to run these tests at this time.
3DMark03 Vertex Shader (FPS)
640x480
800x600
1024x768
1280x1024
1600x1200
6800 Ultra
30.1
29.7
29.4
28.9
28.1
5950 Ultra
19.5
19.0
18.4
17.2
16.2
5900 Ultra
19.3
19.0
18.7
18.0
17.5
5800 Ultra
20.0
19.5
18.7
16.5
16.2
6800 Ultra % Difference
640x480
800x600
1024x768
1280x1024
1600x1200
5950 Ultra
54.4%
56.3%
59.8%
68.0%
73.5%
5900 Ultra
56.0%
56.3%
57.2%
60.6%
60.6%
5800 Ultra
50.5%
52.3%
57.2%
75.2%
73.5%
% Difference from low to high res.
6800 Ultra
5950 Ultra
5900 Ultra
5800 Ultra
3DMark03 VS
-6.6%
-16.9%
-9.3%
-19.0%
Running through the 3DMark03 Vertex Shader test at various resolutions we can see the drop in performance from low to high resolutions. While that performance drop will come from the fact that more pixels are being drawn, the increased bandwidth utilisation may also have performance ramifications for fetching of vertex data that is stored in the onboard RAM of the graphics card. Curiously, what we see here is that despite 6800 Ultra having a greater FPS - more pixels are being written hence the bandwidth utilisation is better - it loses less performance at high resolution than the 5950 does.