LOSTCIRCUITS

Bottlenecks vs. Paradoxes

One thing easily overlooked in light of the established awsome power of the test system used is the fact that there are still system bottlenecks that will cap the performance of any graphics adapter. In other words, we know better than trying to use a 600 MHz Duron for our benchmarks, in that case, most likely all graphics cards used would score about the same. Still, especially at lower resolutions, the CPU and overall system performance is the primary bottleneck with respect to achieving scalable benchmark results. The overall scores of 3DMark2001SE are a good example but a look at the individual scores will tell a different story altogether.

No Antialiasing

For an easy side by side comparison, we have copied and pasted the 3DMark2001SE results before and after (pink background) into the same "Details" pane. Predictably, the high detail benchmarks that are more CPU and memory dependent than the low detail benches do not scale well between the two versions, on the other hand, they don't need to. The only thing to establish is that some of the results will show a linear performance increase of more than 9.5% (which is the difference in core speed) to make our claim that, in fact, the additional pipeline set was enabled.

1024 x 768 x 32 bpp

1280 x 1024 x 32 bpp

1600 x 1200 x 32 bpp

Overall, for most of the benchmarks result details shown, the evidence is marginal, however, the Advanced Pixel Shaders show that at least one of the benchmarks benefits by almost 30%, which is what one would expect if the number of pixel pipelines is increased from 12 to 16. At higher resolutions, bandwidth and fillrate limitations cap the performance increase to 28%.

Next Page:    => 4x AntiAliasing =>