There are several other reasons for the 4GB limit:
o Historically, motherboards were limited by RAM VRM
---- lots of RAM sockets = lots of DIMMs = lots of watts = 30W+
---- RAM VRMs are limited - indeed BTX tackles their cooling
o Historically, memory is relatively expensive
---- PC industry focuses products around usage segments
---- multi-GB buyers tended to buy servers, not general PCs

As always, the computer industry migrates downwards.

Applications using multi-GB were once small & niche:
o dBase servers = RAM determinant
---- more RAM = more index keys, more caching
---- contrast with slow electromechanical HD rotation & seek
---- however such PCs are 1) Servers 2) 64-bit
o Photoshop production machines = RAM determinant
---- historically most pre-press is done on Apples
---- recently reinforced by the Dual G5 Apple re OS-X + 8GB+
---- Photoshop on the PC has deteriorating benefit of 2GB

Remember the focus of the IT industry:
o Create as many interfaces with buyers as possible per unit time
o Eg, CD, DVD, DVD-II speed & iteration, or mobile phone iterations

So some limits are based on guesswork "no-one will need 640KB"
and also one of creating a build-up of pressure for an obsolescence
replacement cycle as opposed to the migratory upgrade cycle.

So there are inherent assumptions in the PC architecture at work:
o Hardware both "internal" & from plug-in devices needs address space
---- efficiency has not been great here - PCI-Express has high demands
o Assumptions by the O/S come into play
---- XP may be based on NT, but it still makes relatively "old" assumptions

A lot of the assumptions are still valid somewhat:
o Most economic buyers of multi-GB = 64-bit = Dual-Opteron/Xeon64 Servers
o Many multi-GB buyers are after somewhat niche machines
---- Apple Dual G5 -- re Photoshop pre-press
---- downsized *nix box -- re displacing IBM/SUN/SGI boxes
o Eventually multi-GB buyers will migrate down to the commodity PC

It comes down to usage/economics:
o Yes CPUs can gobble memory vastly faster than a HD can move it
---- a typical RAID-10 array can manage 150MB/sec
---- a typical P4 CPU can manage 40x that on RAM to CPU transfers
o In reality very few applications run into that as a bottleneck
---- it's a tail-end of the usage distribution - few buyers, high cost, low margin
---- an error is to assume it remains that way - it never does :-)

Adding lots of memory to PCs has various electrical limitations, RAM VRM,
and various electrical dependencies required to maintain a stable I/O system.
That moves you away from mainstream buyer requirements re what they pay.
--
Dorothy Bradbury
www.dorothybradbury.co.uk for quiet NMB & Panaflo fans