So the Mac Pro FSB operates at 1033 MHz but it accepts 667 MHz RAM. So is there much of a performance difference (in terms of CPU<->RAM communication) between Mac Pro and iMac? Is it really worth getting a FSB that only communicates with the RAM at half the speed?
The Mac Pro's Xeon processors are on a 1333 MHz bus (or, to be technically accurate, a 333 MHz bus that is "quad data rate" producing an effective data transfer rate as if it was 1333 MHz. It has two of these busses, one for each socket. Each socket has a 10.6 GB/s data connection. (1333 MHz * 64-bit) For a total of 21.3 GB/s bandwidth between the two sockets.
The Mac Pro also uses up to quad-channel memory at 667 MHz. (Again, to be technically accurate, it is 333 MHz, "double data rate" memory, which is the same effect as being 667 MHz.) You *MUST* use two DIMMs at a time, which enables dual-channel mode. This means that you have the effective data rate equal to one processor bus. (667 MHz * 64-bit * 2 channels) If you use four or eight DIMMs, you get four channels working, for the same data rate as both sockets.
The iMac (and the MacBooks, and the Mac Mini, for that matter,) have a 667 MHz front side bus (again, effectively; it is really 166 MHz "quad data rate") for 5.3 GB/s, and uses dual-channel 667 MHz memory, for 10.6 GB/s. This means that a Mac Pro with four DIMMs has memory bandwidth EQUAL to its processor bandwidth, while any other Intel Mac with two DIMMs has DOUBLE the memory bandwidth as processor bandwidth, because it has half the memory bandwidth, but one quarter the processor bandwidth. (This really helps the Mini and integrated-graphics iMac, though, as the graphics chip has to share memory bandwidth with the processor.)
And this isn't even getting into the technical aspects of the Mac Pro's memory subsystem, which uses a serial memory architecture that saps some of the memory bandwidth, and increases latency over the direct parallel memory architecture of the other Macs.
(DDR 2 is a 'parallel' memory architecture, memory is transferred in 64-bit chunks. The Mac Pro uses 'Fully Buffered' memory, which are DDR 2 chips running through a 'buffer' that turns that parallel connection into a 10/14-bit serial connection to the chipset. It allows more 'channels' of memory over the same number of wires to the memory sockets.)