Originally Posted by pauldovi
However, because your memory has some error correction and other communications build in, slightly higher than double will maximize your processors bandwidth.
Either 1:1 or 2:1 is ideal. Anything different is going to induce some extra delays.
The FSB and the RAM speed are two separate busses that have to comunicate with each other inside the chipset. If they are running off the same base clock, or some even multiple of it, corresponding clock cycles automatically match up. Think of it as tapping one finger at 2 times a second and the other at 2, or 4; different speeds, but they still match up. Running the RAM at an odd divider, like 2:3, induces extra latencies, since info has to go between two totally unrelated clocks. Try tapping one finger 2 times a second and the other at 3 times a second. It's doable, you just have to think really hard about it. Same with the chipset. It has to somehow take bits comming in at a certain rate and sync them to a totally different rate.
This is what separates the men from the boys when it comes to chipset performance. The high end chipsets are very fast at doing these asyncronous transfers, where cheap ones aren't. Back in the old days of Socket A and 370, chipsets at the time were horrendously bad on asyncronous RAM dividers.
Just test, test, test, and see what works best. Like I said before, even if you break even with RAM performance at 666 vs 533, go with the slower speed. You can then run faster timings and outperform your base performance mark.