Double in emulation mode

Not relevant.

wrong part of forum or stupid/not understandable question?

Where in guide does it suggest that double precision values in emulation mode will be stored in a single precision register? When compiling in emulation mode your kernel is just compiled straight up with gcc on linux or cl on windows, neither of which will demote your doubles to singles unless you force them to.

And yes, double precision is only on compute 1.3, requiring the latest GTX 260/280 cards.