Once we got to a mode where we were only doing computation per request, the raw cost is pretty minimal since there's not much CPU time either way.