I am working on a RBPi with a new imager development. I am using CUDA under the Mathematica kernel located in another computer with the NVidia card. Since I am interested in performing parallelism in small devices, I am having the PBPi collect the data and delegate processing to a remote kernel. This solution has worked fine for me.
The second computer uses Mathematica version 9.01 with an NVIDia Geoforce 630 . It is a weak card, but does the job.
The processing power of the microcontroller is not to the level of image processing. I am assuming this is what you are looking to accomplish. Therefore, consider remote kernel parallelism allowing the RBPi tackles only hardware control and data collection.
I hope my suggestion might work good to you.