[time-nuts] Tbolt Temp Specs
Tom Van Baak
tvb at LeapSecond.com
Thu Jan 27 22:14:35 UTC 2011
> The temperature numbers that members have published
> are very impressive.
> My question is: What does the tbolt use for a temperature
> sensor and is there or has anyone been able to independently
> verify that the stated numbers are accurate? Of course if
> they are working extremely well it may be a moot point.
The TBolt uses a DS1620 in a two step high-res mode. In
this case the actual resolution varies from chip to chip and
maybe even across temperature range but in the units I've
tested seems to be under 1/100th of a degree, typically
about 7 mK. I can give you more details if you wish.
Note that this is resolution and not accuracy. But the good
news is that for a closed system like this accuracy is not
important; just resolution.
As for the milli- and micro-kelvin numbers; well, that is a
Given a series of measurements you can always calculate
an average and standard deviation. In a closed loop like a
GPSDO or a temperature controlled box the average error
will always decrease simply because one computes an
average by dividing by N.
But the readings are bounded by virtue of the closed loop.
You'll find the stdev reaches a nominal value after a bunch
of samples and remains pretty constant from then on. The
mean error, on the other hand, has N as a divisor -- so the
larger N gets the smaller the mean error number becomes.
It turns out it doesn't mean anything.
A more appropriate statistic is simply to quote the rms error
value and leave it at that.
Otherwise you get into a game where one person says their
TBolt is stable to 30 mK per hour. The next person waits a
week and says theirs is much better since it averages down
to 0.1 mK. Not to be outdone, someone else waits a month
and claims one micro-Kelvin stability. Hey, check out this
tombstone: "here lies the body of a time nut who kept his
TBolt stable to ten pico-Kelvin over his lifetime".
Quoting temperature regulation with an "average over time"
is misleading. Just use an rms value.
More information about the time-nuts