[Butcher]

Cube maps are a large part of render to texture operations. And they're very big and expensive. Another thing you might end up doing is doing something like a cloth simulation using the GPU. This can output a relatively large amount of data which is needed on the next frame.

Originally Posted by Artic_Kid

For example, if card A generates a given texture at, say, 90% of the way through its frame, and card B uses that texture sometime after 40% of the way through its frame, then passing the texture between the two cards would work. (Likewise the figures 30% and 80%, respectively, would also work. As would 20% and 70%, etcetera. So long at the two figures differ by at least 50% in the proper direction, it would work.)

I suspect an easy way to implement this would be as follows. Simply have the GPUs use the passed texture last (or a late as possible) in the process. So long as it is used after the 50% point in the frame, then this method will always work. (And that's for the worst case where the texture is passed at the very last moment of the frame. If the texture is passed earlier in the frame, then it can be used earlier in the next frame by the other card.) I suspect this protocol is not so difficult to do

The problem is you'd have to code the game with SLI in mind to do that. Most (currently) are not.
Also it's not neccessarily feasible to do this amount of arrangement. There are certain things you have to do in certain orders for them to work correctly. Rendering last isn't usually an option. E.g. if the game is using HDR then it will have to perform tone mapping and such at the end of the frame also the alpha blended objects are generally draw after opaque which adds more dependencies. And of course any FSAA is always done last. These sorts of dependencies will often push the use of the rendered texture forwards in the frame and make overlap inpossible.