I was going to give you a snarky answer, because one of the preconditions for the paper is that it's CPU vs GPU operating on an in-memory database (so there's n

I was going to give you a snarky answer, because one of the preconditions for the paper is that it's CPU vs GPU operating on an in-memory database (so there's no disk access by definition), but in thinking about it perhaps it's not even CPU bound but cache-miss bound, and you could trade the latter for the former by compressing the data into a trie or something smarter than just a B+tree.