I think that the architecture of GPUs will change pretty radically over the next 4 years or so.
Some companies are working on combined GPU + CPU chips in which case one would expect there to be less of a distinction between system RAM and GPU RAM.
Certainly today even a 4 core Intel CPU is starved for memory bandwidth and can't execute a full 128 bit wide load + store from/to RAM per every core per every clock cycle per every core. That is only going to get worse with CPUs of 8 cores or 4 cores + hyper threading or whatever.
Thus RAM bandwidth needs to expand past the current approximate 4GBy/second out to at least 16GBy/second and beyond just to keep up with moderate multi-core CPU based processing.
High performance GPUs usually need around 50GBy/second to 150GBy/second of VRAM bandwidth for today's top generation of single GPU cards.
CPUs with dual channel memory controllers as is now typical have 2 x 64 bit wide paths to the system RAM for 128 bit wide parallel I/O.
Mid-range GPUs also have 128 bit wide data paths to VRAM, with higher end GPUs having 256, 384, or even 512 bit wide paths to VRAM.
The 'Nehalem' generation of PC CPUs / chipsets will have triple channel and quad channel DDR3 models handling 192 bit wide and 256 bit wide paths to system RAM.
http://en.wikipedia.org/wiki/N...em_(microarchitecture)
This suggests a possible convergence of system RAM and VRAM for mid-range GPUs integrated with mid to high end CPUs over the next 2-3 years concordantly with a possible integration of CPU and GPU either on a single chip or on a coprocessor slot / socket basis (quickpath?).
The other factor is that memory prices fall and memory densities increase. It isn't altogether clear that the average desktop (or even gaming) PC will need more than 4GB-8GB of memory for at least a few years until software catches up with the hardware capabilities and 64 bit applications begin to dominate and common applications that can actually use more than 2GB usefully emerge.
Thus the amount of PC ram may stall while the amount of VRAM still is increasing to the 1GB-2GB / GPU level and beyond. Once you have 2GB-4GB of fast GPU RAM you almost might as well use it as system RAM as well if you have the architecture to support that.
Another factor is that graphic resolutions are increasing with 30" sized LCDs becoming common, and bigger ones following rapidly. This will shift the need for VRAM out to 1GB minimum and 2GB common within the next 12 months. This is simply to permit today's "cartoon quality" graphics. Within another 2-4 years GPU power will keep increasing to the point that whole new levels of image quality become feasible with the ultimate goal being basically photo-realistic 3D in real time. That would need something like 8-16GBy of VRAM at least, and probably with new generations of dynamic compression / 3D interpolation et. al.
Another factor is the limited lifetime of 2D displays; eventually we'll start to have more common true 3D display technology in which case we'll see even more need for 16GBy+ VRAM quantities and GPUs that fully outpace today's CPUs in every way. At such a point it will be irrelevant to talk about VRAM vs system RAM since the GPU will be quite enough to be your CPU and VRAM will be far bigger than you likely need for system RAM.