Theoretical Performance

With 16 pixel pipelines and over 500MHz clock speed, Radeon X800 XT Platinum Edition should have a very large theoretical fill-rate. We'll take a look at the XT and PRO's theoretical rates in comparison to ATI's previous high end offerings:

When we look at the X800 XT PE, we see that due to the number of pipelines enabled, and the clock frequency it operates at, the theoretical pixel fill-rate is over 8 Giga-Pixels per second, which is 46% greater than the 12 pipelined, 475MHz X800 PRO and over 152% greater than the 9800 XT. The Theoretical Geometry rages of between the XT PE and PRO are similar as they both have the same number of Vertex Shaders enabled, hence the rates only vary according to clock speed; in comparison to the 9800 XT the X800 XT PE has nearly a 90% greater theoretical rate thanks to having 50% more units and a higher clock speed. The bandwidth differences between the X800 XT PE and the 9800 XT aren’t as large as the fill-rate differences as GDDR-3 memory is only just ramping in the lower speed ranges at the moment, but there is still over 50% more bandwidth available for the X800 XT.

Looking at the X800 PRO we see that is has a 73% fill-rate boost over the 9800 XT, and the same rate for its theoretical Vertex processing abilities. With a 475MHz memory clock rate the memory bandwidth increase over 9800 XT is about 30%.

For the X800 XT PE, the pixel fill-rate difference to the X800 PRO falls a little shy of its theoretical difference, but this is due to this particular test being very memory bandwidth limited (indeed, the performance differences more closely resemble the memory bandwidth differences) – the same can be said when comparing against the 9800 XT. However, looking at the multi-texture fill-rate test we see performance differences much closer to their theoretical rates as this test isn’t as bandwidth limited.

Looking at the fixed function Geometry tests we can see that the performance difference of the X800 XT PE to the X800 PRO is slightly greater than its theoretical rates would suggest, and the difference to the 9800 XT is lower on the single light test but about at the theoretical difference on the 8 light test. Curiously with the Vertex Shader test the performance of the X800 XT PE is slightly lower than the X800 PRO or the 9800 XT, which is likely to be indicative of some other limitation.

When looking at the X800 PRO we see that it nearly reaches its theoretical fill-rate difference to the 9800 XT with the multi-texturing test, but, again, not in the single texture test due to bandwidth limitations. The theoretical geometry rate differences between the X800 PRO and the 9800 XT are nearly achieved in the 8 light test, but again the Vertex Shader test shows a slightly lower performance for the X800 board than the 9800 XT.

Due to the high bandwidth nature of the 3DMark2001SE fill-rate test, the X800 boards both look like they may be configured with half the pixel pipelines, but with double the texture rate – even though this doesn’t make any sense of the X800 PRO configuration (6x2 would mean it was operating on one and a half quads!) we’ll check the fill-rate performances in another test, in this case MDolenc’s fill-rate tester:

As you can see, the “Pure Fill-Rate” performances, whilst still not being very close to their theoretical rates, are higher than an 8x2 or 6x2 configuration for the X800 XT PE and X800 PRO respectively could achieve, clearly marking these as 16x1 and 12x1 configurations. Again, these are the only configurations that make sense as the chips are operating on quads of pixel pipelines.

You may note that the Z fill-rate gets much closer to its theoretical performance, due to this requiring less bandwidth, however it certainly isn’t greater than the theoretical single texturing fill-rate highlighting that there isn’t an optimized Z/Stencil path in R420.