Then at a lower level and smaller latencies it's often interrupt moderation that must be disabled. Conceptually similar idea to the Nagle algo - coalesce overheads by waiting, but on the receiving end in hardware.