3DMark03


5800 Ultra 56.4 45.8 34.8 25.2 19.4
5900 Ultra 57.1 47.7 37.1 27.2 21.1
FPS 0.7 1.9 2.3 2.0 1.7
% 1% 4% 7% 8% 9%

The "Battle of Proxycon" game test in 3DMark03 also makes plenty of use of PS1.4 and again we can see that the 5900 Ultra has a performance advantage over the 5800 Ultra. This test also makes extensive use of Stencil shadows, and from the Fablemark test later on in this review it would seem that some of the performance here may be coming from an increased stencil performance on the 5900 Ultra (but not from UltraShadow, as this requires developer support).

5800 Ultra 34.2 28.1 22.1 16.8 13.3
5900 Ultra 29.0 23.8 18.7 14.2 11.3
FPS -5.2 -4.3 -3.4 -2.6 -2.0
% -15% -15% -15% -15% -15%

The "Mother Nature" game test utilises both PS1.4 and PS2.0 Pixel Shader functionality, and here we can see that the 5900 Ultra has a 15% performance deficit to the 5800 Ultra.

Again, later Pixel Shading tests shows that this is somewhat reflective of other Pixel Shader 2.0 testing due to 5800 operating in FP16 mode and 5900 in FP32 and taking a larger register performance hit because of this.

5800 Ultra 23.9 21.0 17.4 13.6 11.0
5900 Ultra 29.9 25.3 20.3 15.4 12.3
FPS 6.0 4.3 2.9 1.8 1.3
% 25% 20% 17% 13% 12%

With this test we can see that the 5900 Ultra performance is higher than that of the 5800 Ultra. However, oddly, the performance difference is higher at low resolution, rather than high, which belies this being due to increased pixel shader performance as the performance gap would be consistent across all resolutions, or a higher performance gap at the high resolutions.

Now, the fill-rate graph shows that this test is never truly fill-rate (Pixel Shader) limited, but is still quite reliant on geometry performance. The geometry for the scene is buffered in local ram so the performance increase could be occurring because of the increased bandwidth of 5900 over 5800 and hence geometry data is passed to the core faster with the 5900.