But that's just one instruction, highly paralelisable (I assume), as opposed to going through the kernel... Anyways, I haven't measured it, it's just an idea I had!