Okay here we are, the crux of the matter, here’s were both cards will be used for most and rightfully so, if you spend up to $400 on a videocard you expect the best looking image at playable frame rates. Here’s also were we’ll see big differences between the cards and were questions are raised what causes these differences. We already hinted at one probable cause and that’s memory bandwidth, the other two obviously are driver optimizations and the way the cards handle AA and AF processing internally. Upon looking at the scores it is clear that up to 1280x960x32 resolution both cards are more than fast enough with frame rates faster than 60-fps.

At all resolutions, but especially at the highest, 1600x1200x32, we clearly see that the GeForce FX 5800 Ultra is not able to keep pace with the Radeon 9700 Pro. Already at 1024x786x32 the difference is 21% in the advantage of ATi, that’s a rather large performance gap, and one that’ll be hard to close. Although Nvidia has shown that they can work miracles with their drivers and have managed to pull off 10…20% increase in performance as their drivers mature, this may not be enough to give the GeForce FX the performance lead. And to make matters even worse the difference at 1600x1200x32 is nearly 40%, which is a gap you cannot simply close with better drivers. Furthermore the Radeon 9700 Pro has been on the market for several months already and ATi is about to release a new card based on the R350 core, which will be faster than the current Radeon 9700 Pro, but will use the same excellent and mature driver set. Whatever Nvidia does to improve upon the GeForce FX it won’t be enough to best that card.

What we read in between the lines of these benchmarks is that memory bandwidth again roars its ugly head. What’s happening here is very reminiscent of the original GeForce launch and the subsequent move to DDR memory to boost its performance, resulting in almost double the performance. Here we see the 256-bit memory bus of the ATi Radeon 9700 Pro show that memory bandwidth is what counts. Of course Nvidia did not opt for a 128-bit memory bus lightly; I’m sure they weighed the pros and cons carefully and decided that a combination of a higher clockspeed, a new type of memory, and a smaller bus would yield them about the same bandwidth, unfortunately it looks like this is not the case. It is unfortunate to see so much performance in the GeForce FX’ core goes untapped because there’s not enough memory bandwidth, as shown by our little experiment with increasing the memory clockspeed.