Yep, that was a rather bad typo. 15,825 us. No excuse, except I shouldn't post in the wee hours of the morning ;-) I'm not really arguing about measured latency but you have to apply a best-case cost as you cannot measure certain parts of hidden infrastructure than you can't touch. You do know 1000 bits on a 1Mbps line will take "at least" 1000 microseconds.

Even your kernel bypass driver on a SolarFlare still takes at least 40us on 4000 bits from a 100Mbps KRX UDP feed in those older days, as that is simply the best case time on the wire.

True on nanoseconds being the common measure today.