Virtually any card made by nVidia in the last 5 years can be used to run CUDA programs. In terms of “Best”, you’d probably want a Tesla but they are hideously expensive. By hideously expensive, I mean that they are designed to operate only as computing on GPU hardware and not as regular graphics cards. You mount them in servers and “Do Science” to things with them. You can technically mount one in a workstation but it’s still expensive. A bare minimum machine from Dell, for example, is just under $5000.

As far as home applications, whatever card you’ve got would probably work fine provided that it’s on this list. Pick something you can afford. Find some problems that are of interest and amenable to the SIMD(SPMD) architecture and have at it. Finding a reference for C is recommended as well as a CUDA guide. If you don’t yet know any C, Zed Shaw will cure you of that.