Comment by zozbot234

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.