[COLOR=#1951B9]The idea is simple: 3D graphics is a highly parallelizable application. Current generation GPUs all consist of variable numbers of parallel computational units anyway. Why not link them together in a way to take advantage of more cards? At the purely technical "gee whiz" level, this is a pretty cool idea. In fact, Nvidia has even built support for this into their NV40 GPU die. A small percentage of the die is allocated to a high speed I/O port that can be used to communicate between other NV40 GPUs. [/COLOR]
Then of course you have the inevitable drawbacks starting with double the price (for two cards obviously) but more importantly, the need for a platform to support two PCIe graphics cards... overall however NVIDIA has made a step in the right direction, either if you want to look at this as a way to build the latest gaming machine or a great alternative for future GPU upgrades. More here.