I'm no CUDA expert but from what I recall you do need to use different types of memory to write efficient code, or your code might end up being very slow. And

I'm no CUDA expert but from what I recall you do need to use different types of memory to write efficient code, or your code might end up being very slow. And then you have different card models, some share RAM, some don't. Some can process AND transfer data, some can't. Some can transfer only one way while processing. All of this must be accounted while programming.