Well, with the early 3dfx cards you had to daisy-chain the vga cable from the 2d card, to one accelerator, to (optionally) a second accelerator, and the accelerators would switch from pass-through to their output based on the timing of the signal.. This, by the sounds of it, is using the new standard memory management interface to ship data between chipsets internally.
Not *too* different a concept, but a but more flexible, perhaps..
The interfaces are also different this time around, since the AGP era it was made easier to share data between the bus and main memory. PCI-E having more bandwidth and supporting more channels (and even AGP itself) are much better suited than chaining together the cards, as the joint point would actually be the system bus (faster than memory). I do agree that this implementation is much refined, I guess 14 years wouldn't have gone in vain