I have a way simpler explanation. IEEE 754 double can only represent integers up to 2^53 without precision loss, so if you naively average two numbers greater than 2^52, you get an erroneous result.
It just so happens that 2^52 nanoseconds is a little bit over 52 days.
I've seen the same thing with AMD CPUs where they hang after ~1042 days which is 2^53 10-nanosecond intervals.
It just so happens that 2^52 nanoseconds is a little bit over 52 days.
I've seen the same thing with AMD CPUs where they hang after ~1042 days which is 2^53 10-nanosecond intervals.