NVIDIA released CUDA 6.5 this week and it brought support for 64-bit ARM, cuFFT device callback support, CUDA Fortran tools support, CUDA Occupancy Calculator APIs, double-precision performance improvements, static CUDA library versions on Linux, multi-process service performance improvements, and other work.
Thomas Dinges has now shared with Blender developers that CUDA 6.5 is much better off than CUDA 6.0. A few seconds are shaved off the render time while the Volume / Volume Scatter performance is now roughly the same as the stock mode. Checkout that thread for those into Blender and are using a NVIDIA GPU with the proprietary driver. Dropped from Blender is also pre-6.0 CUDA support with 6.5 now being the officially supported version.