well, that problem can be solved...

lucid simply has to make a mini bench for each supported game(or certain effects etc.), and let it run through each installed graphics card (as long as the game is installed... saves time); databases would ignore OC, so thats a no-go

the bench can determine, which card gets which parts of the expected renderings (frames) (ideally those in which its faster than the other installed cards or which are actually supported by the card. like a dx9 card getting all dx9 effects in a dx11 game, as long as the dx11 card cant handle parts of the dx9 code faster, by itself), packing it all inside some kind of log.

then, every single frame has to be analyzed and split up onto the cards, rendered and put together again by hydra.
quite easy doing that with a small scene and with two identical cards (like in the demo. i guess they split up the scene manually, though i hope they didnt), but executing what i described would be quite a bit more demanding

so doing it right is massive work (for the chip and the developers), but with a little game, OS and hardware developers help ... possible