It should be the way of the future. Bandwidth (EFFECTIVE, not actually PHYSICAL) has been the constraint of 3D graphics for a very long time.
R300 and NV25 finally seem to have efficient enough bandwidth optimisation that the effective bandwidth is just about high enough to max out the cores... but not quite. (pure core OC still gives perf. boost)
The idea with an embedded buffer is that you can have a gargantuan path to the memory controller - a mind-numbing 2,560-bit pipe in PS2, and an almost as staggering 1,024-bit (512 each on two buffers) in GCN.
By comparison, the newest PC graphics cards have just started to have 256-bit memory buses - 1/10th that of the GS's embedded buffer.
It makes even more sense in a console environment, since the coder can optimise the cache work very closely, and the frame buffer will never exceed any standard TV resolutions (at least until new standards appear
), so a fixed-size cache can be used without too much deficit.