[keithwalton]

erm some clarifications are in order me thinks.

Agp was a Pentium II thing. it was not in 486's nor pentium / mmx's. It was later introduced into the super socket 7 line for the amd k6/k6-II/k6-III line of cpu's.

Agp 3.0 is backwards complient with 2.0 but not 1.0. and 2.0 is backwards complient with 1.0 It's a voltage thing, agp 1.0 was 5v primary power, 2.0 is 3.3v and 3.0 is 1.5v there are 'humps' in cards / slots to stop you from plugging in incompatible devices.

Also the new PCI standard, PCI express, is not the same as PCI-X, the latter is a 64bit verson of pci and is parallel in terms of architecture. Express or PCI-E x16. Is new fangly and serial. Its twice as fast as agp in the cpu -> gpu direction, and 16x the speed in gpu -> cpu direction.
It is defined in terms of lanes. With x16 having 16 lanes. There is also x1 which is to replace pci 2.3 for less bandwidth hungry boards.

Chipsets have a limit to the number of lanes that they support. I believe the i915/925 chipset supports 20 lanes. which is used in 1 x16 slot and 4 x1 slots. I dont know how many lanes the e7225 (tumwater) chipset has for the xeon platform but it must be alot as it has 2 x16 slots (helpfull for nvidia sli) and several x1 slots.
The upcoming nforce 4 chipset has aledgidly 32 lanes which will also be dished out to support there sli technology. Rumour has it there will be an x8 slot for the second card.


As for graphics peformance, the first batch of pci-e cards are not any quicker clock for clock than there agp equivelents. (pci-e cards may be clocked slightly higher to give them a generic edge)
There is a reason for this, nvidia are using an on die bridge which translates pci-e to agp commands. Since the chip is natively an agp gpu it makes no use of the new features and is infact slower than a native agp card of the same spec in benchmarks sometimes it will equal perf when something else is the bottleneck (most of the time).

As for ati the r42x (420 being agp and 423 being pci-e) have a modular interface design.They are basically identical apart from one small part of the die which is the communcations protocol bit. The bulk of the chip communicates with this section which is either agp compliant or pci-e.

Neither 'solution' specifically demands or even makes use of the new features brought along with pci-e. This reminds me of the whole pci -> agp move several years back early agp cards were basically pci designed cards that sat on the agp bus. It took till the second generation (Agp 2x) cards before any real advantage was seen.

Unfortunutly direct comparisons are hard to make as pci-e boards carry other new features such as ddr2. So performance differences cant be singled out as there are severable variables.

I would though agree with the wait and see attitude. LGA775 is new and it'll take time for potential problems to become more apparent and be resolved (motherboard socket bit is fragile so far). Chips designed specifically for pci-e are on the way by the end of the year and should be a fair bit faster.

K

Answering other points, AGP 1.0 is 1x only. 2.0 covers 2x and 4x. 3.0 is 8x. Also 8x cards do support other os's than xp. 2000 for sure is suported and so is Me (whoever would willingly use that i dont know) 98 support is sketchy. Linux is also supported by nvidia cards :p