I have just realize that ZeroCopy is possible with my GTX460 (use simpleZeroCopy).
Then, I have make a simple exercie to perform Batched FFT (50 R2C FFT 1D with 16384 real simple precision) wo any Host <-> Device transfert.
It is clear that the performance have increased by a factor 100, passing from ~33usec/FTT to 0.37usec/FFT.
I wander if I am wrong as in the CUFFT Library August 2010 there no mention of ZeroCopy capability usage ?
Does someone else is using ZeroCopy in this FFT context ?