Games Tests

Game Test 1: "Wings of Fury"

See here for details of this test.

As mentioned in the 3DMark03 Introduction this test uses a single textured sky, quad textured planes, particle effects along with Vertex Shader Geometry processing.

As we'd expect the 9700 PR0 and 9700 boards are in the lead, however in the higher resolutions all the other boards have dropped back by a higher margin. In the case of the 9500 PRO this must be due to a bandwidth limitation, as it has the same fill-rate as the 9700. The GeForce 4 Ti 4600 is next, and it would seem that its higher bandwidth and greater multi-texturing fill-rate isn't as useful in this test as the higher single texturing fill-rate of the 9500 PRO, ignoring any Z-Cull difference between the two boards.

Matrox's Parhelia is following along very close to the Radeon 9000 PRO, with the latter board even outperforming it at the highest resolution. The Parhelia's quad texturing pipeline is not a particularly good fit for this test and judging by the configurations of ATI's current DX9 boards and NVIDIA's FX range, probably not future titles either. Even though Parhelia has a massive bandwidth advantage over all of these boards, bar the 9700's, its lack of memory controlling optimisations probably means that its not making effective use of it.

Game Test 2: "Battle of Proxycon"

See here for details of this test.

Here we can see that all four of the DX9 class ATI boards have an advantage over the rest of them, even the vanilla 9500 is performing above the GeForce 4 Ti 4600. There's likely to be several reasons for this, the first is that they all are Pixel Shader 2.0 parts, meaning they will have PS1.4 capabilities and thus are able to render the scene in fewer passes than the other boards, which will save on memory bandwidth and vertex processing. Another reason they are likely to be able to perform better in this test is because all the DX9 boards have double sided stencil capabilities meaning that the rendering to the stencil buffer is performed much more optimally.

After the DX9 boards we see the GeForce4 Ti 4600, which narrows the gap to the 9500 by the high resolutions. Studying the difference in the fill-rate graphs of the 9500 and the Ti 4600 we can see that at the lower resolutions the 9500 has a relatively steep rise, before reaching a plateau at 1024x768, which indicates that it is reaching a nigh on pure fill-rate limitation by 1024x768. The Ti 4600 has a steady rise upwards, throughout all the resolutions, indicating that there is some vertex limitations within the benchmark at times, however the FPS is dropping across each resolution meaning that it is fill-rate / Pixel Shader limited as well. Bearing in mind the lack of double sided stencil operations and the need to render over the more arduous PS1.1 rendering path the performance of the Ti 4600 is looking impressive next to the 9500 -- the slightly higher pixel fill-rate, much higher texel fill-rate and probably a more optimal memory bus would be contributing to this; the 9500 also has no Hierarchical Z-Buffer, unlike the rest of the R300 based boards.

Despite the inclusion of an easier rendering path for PS1.4 shaders, which the Radeon 9000 supports, it is still below the rest of the boards, bar the Parhelia. The Parhelia's high multi-texturing fill-rate, although of use during the colour rendering, will be of no use when the stencil buffer is being populated, as this is a pixel fill-rate intensive operation; couple that with the need to do the more strenuous PS1.1 rendering path and it general lack of memory optimisations or Z-Culling routines and its performance is looking less than stellar. During our Parhelia review we made note that we had reservations over Parhelia's configuration and its applicability to future titles such as Doom III, given that this test features a number of similar rendering techniques it seems our reservations were correct.