The mobile market where small,light, and battery life are the bread&butter is where this is going to be a killer feature IMO. Having purevideo and SM3 support in such a low-end part isn't exactly a bad thing the way I see it either.
It is not like the AGP aperture with just the faster PCIe interface either, they have a new MMU, pipeline mods that help negate the effects of increased latency that are caused by accessing the system ram over the PCIe bus, and turbocache management in the drivers which means as the drivers mature performance boosts will almost certainly be seen . I don't think that this would work with high-end cards because the memory on those is too fast for the turbocache scheme to work. It seems that 16mb of system ram@350mhz is the present minimum criteria for it, but that it can allocate enough to make 128mb of total ram available to the GPU. They claim that due to the bidirectional PCIe bus they can get 10.2 GBps effective bandwidth and that thanks to the MMU, pipeline tweaks, and intelligent alogorithms that it can dynamically load balance the system bandwidth between discrete and system ram during rendering from app to app while always keeping the scanout buffer for display refreshes in the cards' ram. the above is from the tech notes, with my speculation added of course.
This seems to be strictly designed for low-end parts, and the entire reason for it is to be capable of using the system ram to augment the cards' in a way that doesn't have the negative impact on graphics performance it has always had in the past. From what little I've seen so far it appears they are succeeding too.
