Ragdoll

See here for details of this test.

The point of the Ragdoll test it to check the performance of both the Vertex Shaders and CPU performance, with the CPU having to do lots of physics calculation and the Vertex Shaders needing to do lots of skinning work. In this instance though, the CPU remains a constant.

The four ATI DX9 units actually have quite strong Vertex Processors, and the 3GHz P4 is one of the highest performance consumer processors available today, so rather than being a pure CPU and Vertex Shader test it is already showing fill-rate limitations, especially on the 9500. All the DX9 units also make use of the PS1.4 path, which will cut down the number of passes made, also cutting down the overall vertex processing.

The other boards all show much more of a flatter rise on the fill-rate graph, indicating that they are much more Vertex limited. Both the GeForce 4 Ti's do have a slight curve downwards at the highest resolution, which indicates that it's getting slightly fill-rate limited. Remember that both the Ti's and the Parhelia will have to render the more arduous PS1.1 code path, meaning they are going to be more Vertex bound due to the increased levels of skinning required.

Despite the Radeon 9000 rendering across the PS1.4 path, lowering its Vertex workload, it still comes in at a little below the Ti 4200 because it has the lowest overall hardware Vertex processing abilities. However, despite, on paper, the Parhelia having a strong Vertex Shader engine it remains at the bottom of the pile in this test, even below the 9000, indicating that either the Vertex Engine isn't as capable, its triangle setup bound, or there is something inefficient in the drivers.

Clearly we can see that the DX9 class hardware here has the upper hand in the performance tests in most cases. Although all the DX9 boards used here are ATI boards, the Radeon 9000 PRO is showing a performance in line with the rest of the DX8/DX8.1 class hardware. In the tests we've briefly discussed the areas why these DX9 boards may have the upper hand than the DX8 boards, despite some of them appearing inferior to other boards in terms of pure fill-rate. It should also be noted that the DX9 boards all have more developed memory optimisations from various compression techniques to Hierarchical-Z buffers that can assist in the performance.

The most notable performance disappointment has to be the Parhelia. Its performance is incredibly low in comparison to its quoted theoretical specifications. While there may be some debate as to whether 3DMark03 is actually representative of current gaming titles out there, there no question that it uses many of the rendering techniques that are being sold to developers by the IHV's and thus many of them will turn up in gaming titles, this being the case one would question the overall longevity of this board.

The tests in 3DMark03 do appear to put DX8 hardware at a bit of a disadvantage, with very low frame rates in many cases, however this has often been the case with new releases of the benchmark and it is designed to be a more forward looking benchmark. The advancement of the 3D industry meant that 3DMark2001 quickly ended up being more limited by the performance of the host system rather than the graphics processor; the onus on this benchmark was to move as much processing across to the 3D board in an effort to help alleviate this system dependency somewhat. The question is, have they succeed in doing this...?

• Please feel free to comment on this article here.