...This is how SLI has worked in times past, however it might have been changed in recent driver releases. I haven't heard anything about it being altered, though, so that is what *I* would assume...
Unfortunately, the behavior changed with the 178-series drivers. Neither card is forced to the other any longer, in either shader cores or clock speeds. Unless you're using Precision, RivaTuner, System Tools or any of the other software overclocking utilities, each card will run at its BIOS-assigned settings.
These changes were implemented not only for support of some of the "Big Bang" goodies, but to make it easier to incorporate the various models of GPUs that would be introduced once nVidia eased considerably on sticking to reference designs with their distribution partners.
ChrisRay, a member of the nVidia Focus Group and one of the Mods at the nVidia Hardware Forum's SLI Zone (as well as their benchmark guru), focused on this question when the 216-cores made their debut in his analysis of SLI with mixed-core GTX 260s
Reduced to its salient points:
Placement of the cards is irrelevant save for one consideration: PhysX support. The added cores of the 216 mean a slight advantage in PhysX calculations. If you intend to run full SLI, then as long as it's enabled PhysX will be handled by the card in the primary slot. Keeping the stronger card there will help in the timing, as the rendering engine must wait on the GPU calculations in this case because the GPU has to switch contexts back-and-forth between rendering and co-processor modes. The stronger card would keep those timing requirements a bit tighter.
Unless you have a completely different card to which you can move PhysX, the only way to get it off the primary card and totally separate rendering from co-processing is to disable SLI. The timing requirements for CUDA support (which provides the PhysX) aren't near as restrictive now because the second GPU is completely designated for co-processing support while the primary does nothing but render, and PhysX doesn't directly affect the timing of the display driver.
This can come in handy for things like Batman: Arkham Asylum, where a GTX 260 can pretty much handle this DX-9 game by itself and you get much better PhysX support and smoother framerates - even in Tri- or Quad-SLI - with a dedicated co-processor. But you don't have to open up the NVCPL each and every time if you set the game's 3D Profile to run Single-GPU mode when SLI is called. Same thing. And here's where the info on PhysX comes from
edited to correct link...