Benchmarks - Geometry Performance
Here we'll use RightMark D3D to look at the performance of the various geometry shader profiles.
| D3D RightMark (FPS) | T&L | VS1.1 | VS2.0 | VS2.0 Static FC | VS2.x Dynamic FC | VS3.0 Dynamic FC |
| 6600 GT AGP | 65.3 | 28.6 | 28.8 | 17.4 | 22.1 | 22.1 |
| 6600 GT PCIe | 65.5 | 28.7 | 28.9 | 17.4 | 22.0 | 22.1 |
| 5700 Ultra | 76.0 | 30.8 | 31.1 | 11.6 | 14.6 | |
| 5800 Ultra | 78.7 | 32.5 | 33.0 | 12.1 | 15.5 |
| 6600 GT AGP % Faster Than: | T&L | VS1.1 | VS2.0 | VS2.0 Static FC | VS2.x Dynamic FC | VS3.0 Dynamic FC |
| 6600 GT PCIe | -0.3% | -0.4% | -0.1% | 0.4% | 0.3% | 0.1% |
| 5700 Ultra | -14.1% | -7.2% | -7.2% | 50.7% | 51.6% | |
| 5800 Ultra | -17.0% | -12.2% | -12.6% | 43.5% | 42.6% |
We can see that all these boards have a much greater capability on whenrendering Fixed Function T&L - the FX boards even more so than the 6600's as theFX boards have some dedicated hardware to assist with fixed function operations,while the 6600's are relying on optimised vertex programs. From there we can seethat the straight VS1.1 and VS2.0 performances get a little closer, but the6600's are much more capable when any kind of branching is used. We can also seethat the performance of the 6600's with static branching is lower than that ofdynamic branching.
Pixel Shader Performance
Again, using RightMark D3D we'll take a look at some of the Pixel Shaderperformances of the AGP 6600 GT alongside the other boards:
| Rightmark3D, 1024x768 (FPS) | 6600 GT AGP | 6600 GT PCIe | 5700 Ultra | 5800 Ultra |
| PS1.1 Procedural | 501.6 | 510.9 | 244.7 | 316.3 |
| PS1.4 Procedural Procedural | 341.4 | 356.5 | 166.8 | 266.4 |
| PS2.0 Procedural | 266.8 | 266.9 | 88.8 | 99.7 |
| PS2.0 1 Light (FP) | 225.0 | 230.8 | 33.6 | 56.6 |
| PS2.0 1 Light (PP) | 299.9 | 320.1 | 43.3 | 66.9 |
| PS2.0 3 Lights (FP) | 124.5 | 127.3 | 13.1 | 24.7 |
| PS2.0 3 Lights (PP) | 166.2 | 169.7 | 20.9 | 34.8 |
| PS2.0 3 Lights (FP) | 54.8 | 55.8 | 4.1 | 8.5 |
| PS2.0 3 Lights (PP) | 74.1 | 74.3 | 9.6 | 11.9 |
| 6600 GT AGP % Faster Than: | 6600 GT PCIe | 5700 Ultra | 5800 Ultra |
| PS1.1 Procedural | -1.8% | 105.0% | 58.6% |
| PS1.4 Procedural Procedural | -4.3% | 104.6% | 28.2% |
| PS2.0 Procedural | 0.0% | 200.6% | 167.5% |
| PS2.0 1 Light (FP) | -2.5% | 569.8% | 297.5% |
| PS2.0 1 Light (PP) | -6.3% | 593.1% | 348.4% |
| PS2.0 3 Lights (FP) | -2.2% | 850.4% | 404.9% |
| PS2.0 3 Lights (PP) | -2.1% | 693.9% | 377.2% |
| PS2.0 3 Lights (FP) | -1.8% | 1237.1% | 544.2% |
| PS2.0 3 Lights (PP) | -0.3% | 674.2% | 524.7% |
Here we can really begin to see the differences between the 6600's and the FXboards. Under integer precision PS1.1 and 1.4 operations the FX boards arebehind the 6600's, however as soon as that moves to PS2.0 floating pointprecision operations there are much more considerable performance differences -this highlights that the FX architecture was tailored more towards fixedfunction and DX8 shader operations operations, whilst the 6600's architecture ismuch more forward looking, but in doing so still increases the DX8 shaderoperation as well. We can also see that both architectures are gettingreasonable gains from using FP16 partial precision operations.
This clearly highlights the difference between the 6600 and the FX's - whilstthe fixed function fill-rate tests had the 5800 Ultra performing very similarlyto the 6600's, there are large differences between them when pixel shaders areput to use, and this is as we would expect given the focus of the GeForce 6architecture, and the arrangement of the pipeline specifically. Given that thefixed function rendering and geometry performances of the 6600's and the 5800Ultra are fairly similar, we'll see how much of a difference the new pixelsshader architecture has one contemporary game titles.
