With unified memory, reading from RAM to GPU compute buffer is not that painful, and you can use partial RAM caching to minimize the impact of other kinds of swapping.