In fact, you do not want to "update the crystal one million times/second".
The whole point of a GPSDO is to combine the excellent short term stability
of the crystal with the excellent long term stability of the GPS signal. If
you update the crystal in real time from the GPS data, you do not need the
crystal...
The control loop of GPSDOs usually have an effective bandwidth measured in
minutes or even hours in the case of rubidium oscillators.

On Wed, Aug 17, 2016 at 11:57 AM, Peter Reilley preilley_454@comcast.net
wrote:

Updating the EFC more quickly reduces the ADEV, though. I find that the fiddly part of tuning a GPSDO design is balancing the ADEV against phase control. If you want keep an iron fist on the phase, you can only do so by constantly swatting around the frequency.

I won't say that getting more frequent phase feedback is a bad thing, but if you're trying to get the PLL time constant to be longer rather than shorter that it won't help a lot.

Sent from my iPhone

Wouldn't you also not be able to actually sync to the individual chips,
since you can't really see the start of any given chip so much as you
just see the correlation over larger sections of the stream? Plus you'd
have to track only one SV at a time (right? Since I doubt the edges of
every chip are perfectly aligned across all SVs even under the best
conditions), so things like brief multipath excursions or even
atmospheric/ionospheric fluctuations would push you off by a bit as well...

(which is why, of course, you have the long control loop that GPSDOs use)

-j

On 2016-08-17 11:41 , Didier Juges wrote:

Hi

You can update the EFC a billion times a second.  Update rate and bandwidth are not the same thing. If you want good ADEV, the loop better not have a bandwidth greater than 0.01 Hz. GPS ADEV is pretty awful at 1 and 10 seconds. It is starts to be good past a few thousand seconds. Yes, older modules are a bit worse than newer ones. Also sawtooth correction can make things a bit better.

Bob

Sent from my iPad

Well, on a practical level, if you update the EFC that frequently then the DAC change glitches will dominate the actual output even if you’re not actually moving the needle much.

Hi,

I agree.

There is however a subtle detail, how they leak out over time.

At one time we had to lock an 155,52 MHz oscillator up to 8 kHz, this
for a 2,48832 Gb/s link, which needs to pass the SDH STM-16 jitter and
wander specifications. The first attempt at that PLL was using a 4046,
and the charge-pump was being used. The charge-pump has dead-time, and
well, they thought it was good to only push the EFC here and there. What
this meant was that they created a triangle-waved frequency modulation
of low rate, which then created phase modulations as it went through the
integration of the oscillator. The scale-up factor made this quite
noticeable at the actual bit-rate. It made the point that you need to
update often to keep deviations limited, and when doing it at a higher
frequency, they are easier to filter out.

In essence, you need to think what each comparison or update creates as
a step response and how it is averaged out over time.

In this regard a PWM is a really bad signal, as it can push the
strongest amplitude at the lowest frequency, which becomes hardest to
filter. For one design I needed to increase the resolution, so I made an
interpolation but with inversed spectral density to that of PWM, to push
the highest amplitude to the highest frequency so that filtering becomes
easier. Turned out to be quite easy and work well.

High update rates can be very useful even if the bandwidth of the loop
is low. The bandwidth only limits how low the updaterate can be, but the
phase-noise purity makes update rates and smoothing mechanisms interesting.

Cheers,
Magnus

On 08/17/2016 11:53 PM, Bob kb8tq wrote:

Hi

A mixer style phase detector running a GHz range oscillator is one example of a
system that technically updated the EFC several billion times a second. There
does not have to be a DAC involved.

The point is still looking at the noise characteristics of the oscillator and the reference.
It is best done in the frequency domain as phase noise. We substitute ADEV, but that
is not an ideal proxy. Either way you want the loop to cross over from one to the other
somewhere in the vicinity of the “equal noise” point if it exists. If there is no equal noise
point, that makes you wonder a bit about why you are locking one to the other :)

Taking ADEV, since that’s what we have the data on:

GPS starts out somewhere in the 20 ns to 0.2 ns range at 1 second depending on what
you are looking at, which module you are using, and who you trust for your data. That
compares to a good OCXO that should be in the 0.001 ns range at 1 second. There is a
long way to go (larger time span) before the OCXO and GPS are anywhere near
the same noise level. You will need to get to 200 seconds with the best GPS numbers above.
You will need to get to 100X that with the worst ones. Yes, there is a bit of hand waving
in all of that.

Bob

Good point, and an example of how good digital filtering (helped with upsampling) can make the design of the analog filter much easier :)

Reference the digital audio battles of the past century when 1 bit D/As running very fast started replacing the expensive 16 bit audio DACs running at 44kHz.

Didier

On August 17, 2016 5:25:39 PM CDT, Magnus Danielson magnus@rubidium.dyndns.org wrote:

--
Sent from my Moto-X wireless tracker while I do other things.

Bob wrote:

Phase noise and xDEV measure the same thing -- the stability of an
oscillator at different time scales.  They just express the result
differently.  Phase noise expresses it as PM in the frequency domain,
and xDEV expresses it as "parts per" in the time domain.  (Yes, this is
a somewhat simplified view of it, but it captures the essential point
without undue complexity.)

Conventionally, we switch from using PN to using xDEV at a time scale
(reciprocal frequency scale) of around 1 second, but there is no
mathematical reason why they both cannot be extended indefinitely in
either direction.  The convention arose largely because the equipment
and techniques we use[ed] to quantify them have traditionally been
different at time scales (reciprocal frequency scales) greater than and
less than about one second.  Now that we are in the era of "digitize
everything, and let Laplace sort it out," we needn't view it as the
rigid convention it once was.

Not really, if one has lower noise at all time (frequency) scales, just
lock to that one at all scales.  (It may call into question why you're
fiddling with two oscillators, rather than just using the output of the
quiet one, if they are both at the same frequency -- but there are a
number of reasons one might want to do that.)

Best regards,

Charles