Hi all,
how many CUDA device the system will see using a GTX690?
If the system sees 2 CUDA devices would it mean that the two GPU do share the
on board Memory (no need to memcpy)?


CUDA will see two independent devices, each one with its own 2GB of memory.
You will still need to memcpy, but using p2p you should be able to use the internal gen3 switch and achieve pretty high transfer rates.

Can the two CUDA devices still communicate at gen3 speeds if plugged into a gen2 motherboard?

Yes, if you are using P2P.


What’s the API interface for that?


The Best Practices Guide says, “UVA is also a necessary precondition for enabling peer-to-peer (P2P) transfer of data…,” but I can’t find any information on UVA in the Programming Guide or the Toolkit Reference. It doesn’t even say that UVA is related to P2P! What am I missing?

you need to be running a 64-bit process in order to enable P2P, basically. also not WDDM.