I'm no motherboard engineer but it seems to me that the nVidia nForce 2 motherboards with onboard GeForce 4 graphics don't work this way
nVidia's solution is a custom chip that has the same capabilities of the GeForce 4, and uses a custom bus controller (some Hypertransport variant) to do memory access. This is why it's so "fast".
What I was talking about was your plain ordinary common-or-garden embedded chipset - like an Intel i845 or VIA CL266 with that "Chrome" graphics. Not a fully fledged GPU, but just a quick fix to get video onto the motherboard - not high end, not really suitable for games (even the i845 plays Quake III poorly.. it's bolstered by the fact that the minimum speed CPU to use with it is some 2.4GHz)
The Intel variants of these chips used to - I don't know if they still do - come with the option of a special "cache" AGP card which was basically faster memory so that it didn't have to shuffle around using slow system RAM (this was in the days of PC133 SDRAM, perhaps DDR & huge FSB speeds have negated the need for it).
As it stands, anyway, you can't get around the fact that you don't need AGP faster than 1x with a decent graphics card. With an 8MB Permedia card, you're going to be in trouble. With a Radeon 8500 at 64MB average memory, you won't.
As for Eric_Z, what on earth are you going to use PCI-X for? I'm telling you that you don't need PCI-X cards in a Pegasos because that's not the market (huge servers and high-performance dual processor workstations, like the PowerMac G5). Most people can't afford PCI-X cards, if they even exist for the purposes people think they want them for, and don't need the added performance.