Wow! That's a good question!
I've been looking for that for a long time and didn't find anything that fully explains the relevance of the different parts of graphics architecture. Those links at least tell some basics, but I fear that they only scratch the surface. I know everything that a person outside Nvidia or Ati needs to know about the technology behind them, but knowing to wich extent they are relevant to performance (I think you are looking for this), is a different story. Basically all of them are equally relevant, but always one of them is going to be the bottleneck, so balance is the answer. A really difficult answer as you can take out from the fact that Ati and Nvidia chose so different paths, and change them constantly. So for every generation and architecture proposition, the part that breaks the balance first, becomes the bottleneck, and thus the more relevant one, the one they must evolve to get better performance.
Also, it is important to note, that the relevance you are refering to, is tied to the use you (or game developers) are going to give to the graphics card. For example memory bandwith gets more relevant once we start to crank up the resolution, AA and AF. Despite this, I would say that today memory bandwith is one of the less relevant things, because it has exceeded it needs (always talking about performance parts, because DDR2 is most times than not insufficient). A proof of this is the path that Nvidia has taken with the 8800GT, and Ati with the HD3800 series. They have nearly half the bandwith and less pixel fillrate, still they perform as well or better.
Until the 8800GT came out and I saw some benchmarks, I used to consider texture fillrate as the more relevant one, with shader power following close. Now I think shaders are the more relevant. That is pretty obvious to me, if you make the math (SPs x shader clock) for the GF8800 series and compare them to the actual performance of the cards on the benchies. They completely match. Still since on GF8 series shaders and TMUs are grouped on "quads" so for more SPs they have proportionally more TMUs, its hard to tell if bottleneck happens on SPs or textures. Since 8800GT arrived I place my bet on SPs, because where G80 had 1 Texture Address Unit for every 2 Texture Mapping Units, the G92 has double the Address units, and there isn't a significant jump on performance compared to G80. I expected this jump, since even if peak fill rate remains the same for every quad, the added addressing power should make the average a lot higher, and thus increase performance if this was the bottleneck.
Summarizing, the relevance of different parts as I see it is as follows:
SPs power > Texture Fill Rate >>> Memory bandwith > Pixel Fillrate* > Memory amount**
* I place it the least relevant because at around 12GP/s (2900XT, 8800 GTX/Ultra) you have enough power to reach as high as 1920x1200 16X AA/16X AF without bottlenecking and I don't see the point of going much higher on the future, but I do know that higher memory bandwith is going to be needed on the future, just not as much as they first thought when they jumped to 512bit or maybe even 384bit. I'm 99,9% sure that they made the jump because of marketing though.
** And yeah, the least relevant is memory quantity, as long as you have a minimum.
That is the World of Graphics Computing Power Relevancies from my point of view. Hope it helps you make an overall idea on how that works, although I wouldn't take it too seriously, because I have been (alongside with TPU, Tom's, Anand... wiki...
yeah, whatever...) my own teacher on this...
See you!
EDIT: When I said "G92 has double the Address units" I meant it has the same number of TMUs and TAUs, compared to half the TAUs versus TMUs on G80.
