Yes, Windows does recognize both cards. Make sure you install your drivers, or at least updated ones if available for your cards. So far with my older GTX 580 Classified and my GTX 780 Ti, the same drivers work for both. Windows doesn’t have a problem with that.
Don’t quote me on this, but I think you can even use a cheap ATi card only for displays, and then have an unconnected-to-displays Nvidia card just for rendering. (Could possibly be conflicts with drivers, be careful).
I haven’t had any problems switching the displays from card to card. Windows just does it, no need to tell any software to do anything. Obviously you’ll get a black screen when you unplug the displays, for me the screen came back on pretty much as soon as I plugged the cable into the other card.
I should also add that with the Nvidia cards at least, when you use multiple cards, through the Nvidia panel you can select one of the cards as the PhysX processor for games or applications that support it, so you can offload some computation from the display card to the secondary card. At least when I do play some games, in some way both cards are being utilized.
You will see all tiles being rendered together, even from the unconnected card. When I render something using Cycles GPU-CUDA setting, with one card I would see only one tile rendering, with both cards I see two tiles rendering, and overall the scene renders much quicker because of this. You do have to set Blender to use both cards, and I believe it is recommended that you DO NOT use SLI for this.
I should also note that when I used to have just the GTX 650 SC (display card) and GTX 580 Classified (render card), that when I used them both together with the BMW Benchmark, it actually took much longer to complete. The 650 slowed everything down because Cycles doesn’t combine GPU from both cards to render a single tile, instead it queues each tile to the availability of either card – in this case the 580 did most of the work because it completed its tiles much faster. And while the 580 finished rendering most of the scene in a few seconds (rendering 3 of the 4 tiles) I still had to wait about 2 minutes for the 650 to finish its one and only tile.
What this means is that your overall render will probably only be as fast as your weakest card can finish rendering its tile(s) if you use mutli-GPU.
My original idea for using a weaker cheaper card wasn’t to increase rendering speed by using it (I already knew it would be too slow), but to run three concurrent displays (and to solve a CUDA error with my 580). It was simply a bonus that offloading my displays from my 580 actually allowed it to render significantly faster since its only dedicated to rendering.
And one more note. While my 580 did see a significant increase in render speeds from the BMW bench (from 48 seconds to 30) when it was no longer being used to display video, this trick didn’t work as well with my 780 (from 47 seconds to 42). So depending on the series of cards you end have, you may or may not see any real increase by offloading its job from displaying video.