If you want to load 1 billion addresses and work on them at the same time (in parallel), there isn't enough device memory to fir everything on current GPUs according to my above calculations. So you'd have to break it into parts and place more copy calls which will affect performance.

Every time NVIDIA increases GPU memory this brings two kinds of speed benefits - one is you can process more stuff on the GPU at the same time, and it will be faster (as CUDA cores usually become faster at the same time) and the other is that as a direct consequence, there's less latency delay from excessive CopyHostToDevice calls and vice versa.