CUDA is a mixed bag. I use it in a Distributed.net project, RC5-72.
It can really slow down your desktop, making it seem like you are working in a cold barrel of molasses. I wouldn't recommend using it unless you have multiple video cards, one of which you will use for CUDA processing. That is how I have my desktop set up. The built-in Nvidia 8300 GPU is controlling my desktop. The add-in GeForce GT220 card does not have a monitor hooked to it. Instead, it is running the CUDA client. Thus, my desktop runs well.
Also, I noticed that with dnetc running, if I go to print something, it has been interfering. I have to halt the CUDA client, print, then resume the client. I think it has to do with the graphics calculations during the printer rasterization prior to sending the print job to the printer.
In any case, the blender client may or may not exhibit these issues. I am telling you just so you will be aware of what may happen if you can get a CUDA blender client running.
Now for the CUDA libraries, the distributed net AMD64/CUDA client came packaged with them. They just needed to be made visible to the system. In my case, I copied them to the appropriate /usr/lib64 area and linked them to /usr/lib. It was that or edit path statements to point to the directory where they were located. What I did was easier for me. You might find the libraries are packaged with the blender client. If not, you might need to download them from Nvidia. The blender site may be your best source of information on how to get CUDA going for that project.