SP made a very detailed post a while back concerning VRMs. I'll post it here in full size text:
==============================
Actually there are some new P4 Xeon boards that use VRM. My P4DC6 uses VRMs as does the P4DCE. The P4DC6+ and P4DCE+ do not have add-on VRMs. The plus models have the VRMs integrated into the board. Thia was done to reduce manufacturing cost and because the add-on VRMs interfered with some heatsinks fitting these boards. However, if you take a look here at Tyans new Grand champion HE Quad Xeon board shown on this page you'll see that it also uses add-on VRMs and has little black slots next to the CPU soockets where the VRMs plug in.
http://www.tyan.com/products/html/thundergche.html
So, as you can see it's not just old boards that use VRMs. The reason you don't see them that often is they are mainly used on high-end boards and not your everyday ordinary dual processor board.
As for the reason why add-on VRMs are used it's because it's thought that since they are power handling components they may be slightly more prone to failure than many of the other components on a motherboard and can thus be replaced if one should fail without having to replace the entire board and that can be significant considering that these type of boards tend to be fairly expensive. Also, removable VRMs do offer the potential for upgrades in some situations. However, the comments about VRMs being used because there was no software voltage adjustment is really incorrect. It really has nothing to do with software voltage adjustment. When a board is made the VRMs are made to produce a range of possible voltages for the CPU based on whatever the current specification is at that time and whatever voltages future processors are expected to run. The CPU tells the VRM what voltage it needs through the VID pins on it and the VRM must be capable of producing that voltage. So, most VRMs, even the add-in can can produce a range of different voltages for both current and future CPUs. However, the problem comes in when specifications change and newer and newer processors are introduced that run at lower and lower voltages. It possible that if a board has a long life span toward the end of it's lifespan newer cpus will come out that were never planned initailly and weren't part of the specification when those VRMs were produced. That means the VRMS may or may not recognise those VID signals and may or may not be able to supply those voltage since at the time they were produced no such voltages were part of the spec. In a case like this software voltage adjustment is irrelevant because all software voltage adjustment lets you do is adjust the voltage over the range the VRMs hardware allows and if the hardware goes no lower then software voltage adjustment is useless. In a situauiton like this being able to remove the VRMs and replace them with newer ones made to a newer spec could extend the life of an expensive motherboard. For an example look at 440BX boards when coppermine came out. Initially these boards were made to run Klamath, deschutes, and Katmai processors and with each of those new processors the voltages changed. Finally when Coppermine came out the voltage had drop all the way down to 1.65v from the 2.0 to 3.3v range of all those earlier processors. Needless to say there were some 440BX boards that just couldn't go any lower because at the time they were built it wasn't thought they would ever need to support a 1.65V CPU. So, many of those boards weren't compatable with coppermines because of this. If they had removable VRMs then there would have been the possibility of upgrading them to support the newer porcessors at least as far as voltage was concerned anyway. So, that's a couple of reasons why VRMs are sometimes used in high-end boards even to this day.
===============================
Thanks, SP!