[time-nuts] NCOCXO anyone?

[David]

On Fri, 22 Jul 2016 11:33:12 +0200, you wrote:

>On Thu, 21 Jul 2016 22:22:14 -0500
>David <davidwhess at gmail.com> wrote:
>
>> Increased integration has only helped insofar as you can attempt
>> designs which would have been prohibitive before.
>> 
>> I keep trying to come up with a charge balancing design but what about
>> Linear Technology's solution from back in 2001?
>> 
>> A Standards Lab Grade 20-Bit DAC with 0.1ppm/°C Drift
>> http://www.linear.com/docs/4177
>
>You can already get 24bit DAC's off the shelf from TI (DAC1282).
>I do not know how stable they are in reality.

There are a lot of DACs like the TI DAC1282 which are not primarily
intended for DC applications.  At least according to its
specifications, it's gain error drift is 100 times greater and its
offset error drift is 150 times greater than the LTC2400 ADC used for
discipline in the Linear Technology application note.  The best DC
DACs I could quickly find are only 4 time better than that so still
more than an order of magnitude below the performance of ADCs.

>I looked into high
>precision DAC's a year or two ago and figured out that once you
>cross the 20bit line, all kind of weird stuff happens that is
>hard or almost impossible to compensate for. The trick with using
>an ADC (which are available up to 32bit) doesn't really work either,
>as offset drifts, thermal voltage etc are hard to impossible to
>compensate completely. If you go through the volt-nuts mailinglist,
>you'll see how much effort it is to even keep 1ppm (~20bit) stability
>of a voltage reference... and then to get that performance out of a DAC.

If you expect analog specifications in line with the claimed digital
resolution of ADCs and DACs, you will be disappointed.  20 bits is
about where they top out no matter how many bits are available; the
best you can hope for is that they are monotonic but how meaningful is
that when it is buried in noise?

The LTC2400 is considered suitable for 6 digit designs before software
calibration is used which the application note and datasheet mention.
In this case, it is its repeatable INL which can be corrected for and
its low gain and offset drift which matter.

It is too bad voltage control of an oscillator cannot be made
ratiometric.  Or can it?  I have never heard of such a thing.  That
would remove some of the demands on a low drift reference.

>If anything, I think the better approach is to use high resolution DAC
>like the DAC1282 or maybe the DAC1280 with a custom modulator and put
>it inside the control loop such that the real measurement happens in
>the frequency/time domain. The results from Sherman & Jördens[1]
>seems to indicate that sub-100fs stability should be possible, though
>there are a couple of open questions there.
>
>			Attila Kinali
>
>[1] "Oscillator metrology with software defined radio",
>by Jeff Sherman and Robert Jördens, 2016
>http://dx.doi.org/10.1063/1.4950898
>(it's available from NIST as well)

Based on Rick's description of the problem, it seemed like that was
what HP tried and it did not work because the DAC had too much drift
to compensate for in the frequency/time domain:

>20 years ago when HP was working on a so called "smart clock"
>GPS box, they decided to do what you said, use a DAC to
>tune the EFC on the E1938A oscillator.  I
>recommended to them NOT to try to do that, but they
>didn't listen to me.  At that time, it
>was nearly impossible to come up with a DAC, buffer
>amplifier, and voltage reference that didn't degrade
>the stability of the E1938A, which isn't even as stable
>as a 10811.  What you need to ask yourself is:  in
>2016, can I finally get analog components that are
>an order of magnitude or two better than what was
>available in 1996?  I don't know, without researching
>it.  Certainly, we can't depend on Moore's law coming
>to the rescue.  If anything, that works against analog
>IC's by obsoleting older analog processes.