64-bit CUdeviceptr

cuda.h defines CUdeviceptr as an unsigned int. Is it possible to allocate more the 4GB using the driver api? Fermi supports 64-bit pointers and there are errors for improperly using 64-bit pointers (CUDA_ERROR_POINTER_IS_64BIT and CUDA_ERROR_SIZE_IS_64BIT). Also, adding integers to a CUdeviceptr seems to produce new valid CUdeviceptrs at the appropriate offset from the original pointer, is this behavior specified in the manual somewhere, or could it change in future implementations? Thanks.