The problem with Gigabyte's method well not problem but limitations of the PCIE chipsets at the moment is say you did put 2 3D1 cards together for a 2Xdual GPUs and it well worked..but it would work with one GPU recieving only 4X PCIe lanes. With SLI on consumer motherboards, when in SLI configuration each slot takes only 8X for a total fo 16X. This is a limitation of the chipsets nForce4 anyway. Only professional chipset like the nForce Pro 2200 version is able to deliver 16X to each card in SLI mode.
Using a onboard controller is quite good idea when you come to think of it, in that the CPU (thru drivers) doesn't have to any work to decide what when where etc (in terms of DATA) ie load balancing each cards gets. However an onboard controller would have to quite proficient in it's capabilties and may drive costs up of the chipset.
SLI is not a fad. SLI is the future IMHO be it NV or ATi's version. They will give performance benefits once all the teething problems (like seen now) get ironed out. Some websites just do not know how to set up profiles. Chrisray has extensibly investigated performance and profiling. It is a question of deciding which method of SLI is best, AFR or SFR.
Also, most of the times, SLI is limited by CPU as well. The GPU is waiting on the CPU to finish things up before progressing, like triangle setup. Once this moves to the GPU aka PPP, this might not be an issue. PPP maybe included in next gen GPUs like nv50 or r520. But it's big "might"..