Anyway, a perhaps separate question is whether having two dice in a single board is really advantageous in any fashion (programming, performance, etc.) over using two different cards.
The physical proximity allows for certain potential benefits, though the current implementations of Crossfire-on-a-stick leave all or most of them on the table.
If the sideport in RV770 were actually used, there would have been an additional link between the chips with some extra bandwidth not available to a two-card solution. AMD stated that there was some benefit, but that the driver team improved the software so much that it was no longer as compelling to force board partners to actually use the side port. (I can't find the links, but I think the biggest benefit was some improved minimum frame rates. Peak benchmark numbers didn't go up all that much so...)
The PLX bridge chip in theory allows for slightly faster communication between the GPUs. There might be some advantage if it allows for full 16x communication independent of the motherboard the card sits in.
If GPUs were architected differently, there could have been savings in the amount of DRAM needed, less CPU overhead, and maybe the GPUs could have appeared to the system at large to be one single GPU, avoiding the drama of required profiles, Catalyst AI, any number of other examples of GPU laxity.
The crossfire-on-a-stick model gives all the headaches of multi-card solutions and driver shenanigans, though it does possibly save a PCIe slot, depending on the cooling solution and motherboard layout.
I suppose that's an advantage, I guess.
