Very good article. Now that makes sense. One card renders one frame and then the other one renders the next. While one renders, one processes. Sort of like a circular sound buffer. While one section is playing, the previous/next is loading. Now that would be a performance gain.
As for the actual SLI mode, again I'm very suspicious of all this branch prediction stuff. So the only way a game can take advantage of SLI is simply to throw all the vertices at the card during loading and then they can choose from that pool. What that article fails to answer is the fact that we are still sending vertices via the bus be it PCI-X or AGP. So if you need to send vertices to the card during the game, it's possible that SLI will actually be slower than non SLI. Since video card memory is no where near enough to stick all your vertices into, plus textures, etc,. I think the other mode where they take turns rendering frames would be the only mode where a performance gain could be attained.
Well, at least I wasn't too far off. I knew that both cards had to have access to all the vertices because prior to transformation there is no way to know where in screen space that vertex will be.
