Games Benchmarks - UT 2004 (DirectX)
The first game benchmark we'll look at is UT2004. Although UT2004 is relatively recently released, to our knowledge it is still largely based on a similar build of engine as UT2003, which means its feature set utilisation still marks it primarily as a DirectX7/8 engined title. We are using a custom Firing Squad benchmark demo.

UT2004 (FPS) | 640x480 | 800x600 | 1024x768 | 1280x960 | 1600x1200 |
6600 GT AGP | 81.2 | 83.2 | 83.0 | 80.3 | 64.2 |
6600 GT PCIe | 68.5 | 69.1 | 68.8 | 68.4 | 64.6 |
5700 Ultra | 91.0 | 86.0 | 65.6 | 46.8 | 33.0 |
5800 Ultra | 87.2 | 87.9 | 84.2 | 71.1 | 51.6 |
6600 GT AGP % Faster Than: | 640x480 | 800x600 | 1024x768 | 1280x960 | 1600x1200 |
6600 GT PCIe | 18.5% | 20.3% | 20.7% | 17.4% | -0.7% |
5700 Ultra | -10.8% | -3.3% | 26.6% | 71.8% | 94.4% |
5800 Ultra | -6.9% | -5.4% | -1.4% | 13.0% | 24.5% |
Seeing as UT2004 is fairly CPU bound, even on modern mid-range graphics boards, we see that the AGP 6600 GT has a higher performance than the PCI Express version in most resolutions, despite its lower bandwidth, due to the platform differences. At 1600x1200 the performances do converge though.
In comparison to the 5700 Ultra the AGP 6600 GT has nearly a 100% performance advantage at high resolution. Looking at the 5800 Ultra we see that even this is falling behind the 6600 GT's by 1280x960, despite it having very similar fill-rate and bandwidth specifications.

UT2004, 4x FSAA + 8x AF (FPS) | 640x480 | 800x600 | 1024x768 | 1280x960 | 1600x1200 |
6600 GT AGP | 81.6 | 80.6 | 65.4 | 48.5 | 24.7 |
6600 GT PCIe | 68.7 | 68.6 | 65.9 | 51.9 | 30.2 |
5700 Ultra | 58.2 | 44.1 | 32.0 | 23.3 | 14.5 |
5800 Ultra | 75.0 | 59.1 | 43.5 | 32.6 | 18.2 |
6600 GT AGP % Faster Than: | 640x480 | 800x600 | 1024x768 | 1280x960 | 1600x1200 |
6600 GT PCIe | 18.9% | 17.5% | -0.7% | -6.5% | -18.5% |
5700 Ultra | 40.2% | 82.7% | 104.4% | 108.6% | 70.5% |
5800 Ultra | 8.8% | 36.5% | 50.4% | 49.0% | 35.6% |
Here we can see that the 6600 GT's are still fairly CPU bound under this UT2004 test, up to 1024x768, and then things become more fill-rate and memory bound. However, at 1600x1200 there is a large fill-rate drop off as the frame-buffer requirements for 4x FSAA at 1600x1200 push some textures across into system RAM, meaning that the have to be addressed across the graphics interface. The PCI Express 6600 GT is affected by this slightly less than the AGP board as it has a higher bandwidth with the PCI Express interface.
The 6600 GT's have again got fairly large performance gains over the FX boards. In the case of the 5700 Ultra some of this performance difference will be dues to the difference in texturing performance. However, given the performance difference between the newer generation boards and the 5800 Ultra, which has the same theoretical texture performance as the 6600 GT's, some of the performance difference may also be due to more effective FSAA compression routines.

6600 GT AGP | 640x480 | 800x600 | 1024x768 | 1280x960 | 1600x1200 |
Normal | 81.2 | 83.2 | 83.0 | 80.3 | 64.2 |
8x AF | 81.6 | 83.6 | 80.2 | 64.4 | 48.1 |
4x FSAA | 81.9 | 75.7 | 57.9 | 32.1 | |
4x FSAA + 8x AF | 81.6 | 80.6 | 65.4 | 48.5 | 24.7 |
% Diff from Normal | 640x480 | 800x600 | 1024x768 | 1280x960 | 1600x1200 |
8x AF | 0.5% | 0.5% | -3.4% | -19.8% | -25.0% |
4x FSAA | -1.5% | -8.8% | -27.9% | -50.1% | |
4x FSAA + 8x AF | 0.5% | -3.1% | -21.2% | -39.6% | -61.6% |
Looking at the detailed performances of the AGP 6600 GT under this UT2004 test we see that there is a maximum performance penalty of 25% for 8x AF (Anisotropic Filtering), and with 4x FSAA prior to 1600x1200, which has the larger performance drop due to textures being pulled across from system RAM, the performance drop if up to 28%.