Hi Mark,

Not to worry. It's not really "too coarse". Your method is more than enough for millisecond or microsecond timing. Consider that everything about time & frequency is merely an exponent; avoid fuzzy words like coarse or fine. Your approach will work just fine; many a GPSDO has been designed that way.

If you measure your oscillator to a second using WWVB and make adjustments, or if you measure your oscillator to a millisecond using NTP and make adjustments, or if you measure your oscillator to a microsecond using GPS and make adjustments, it is really all the same thing. The exponent is different but the concept is the same.

The final result depends on the time quality of the external 1PPS, the frequency quality of your oscillator, the resolution quality of your measurement, and the granularity of your adjustment. So every GPSDO "works", the question is simply how well does it work? And does it meet your needs? For many applications 1 ms or 1 us timing accuracy is more than enough. For only a few applications does 100 ns or 50 ns matter. It gets much, much harder as you get into the double and single ns digit range.

Several members on this list have developed GPSDO that are based on a MCU timestamping of 1PPS pulses. It works fine. Perfecting the tuning constants takes some time, and is somewhat dependent on your GPS receiver, your antenna, your location. Your oscillator, environment, and MCU also play a key part. But the art is to be able to measure all of this, to document it, to experiment, and to possibly improve on it over time.

/tvb

----- Original Message -----
From: "Mark Haun" haunma@keteu.org
To: time-nuts@febo.com
Sent: Wednesday, February 26, 2014 1:51 PM
Subject: [time-nuts] GPSDO with all-digital phase/time measurement?

Hi Mark,

I'm neither an engineer, nor an expert, but here are my comments. 

I think that the idea of 100ns/T is wrong.  There are several variables that control accuracy, but the time between pulses from your OCXO (assuming no phase or frequency drift) isn't one of them.  So, that gives 1/T.  Here the problem is that T must get large before your accuracy can be good.  You can achieve very good accuracy, but at the cost of waiting thousands of seconds between "phase points"; i.e. where your 1PPS coincides with the 10 millionth OCXO pulse.  The theoretical maximum would be infinity, of course, but your oscillator won't be that stable.

Another big problem is the accuracy of the 1PPS pulse.  I'm using an Adafruit GPS receiver, and it's listed as accurate to within 10ns.  And it is, but you have to be wary of exactly what that means.  It doesn't mean +/- 5ns.  So, as your 1PPS pulse bobs back and forth, you will often encounter an OCXO pulse up to 10ns early, or up to 10ns late.  So, might you count 9,999,999 pulses from the OCXO immediately followed by 10,000,001 pulses.  Neither of those, by itself is a signal to change the EFC voltage to your OCXO.  In fact, it is normal for your count to alternate between the two for long periods, if you are very very close to exactly 10MHz, just from the quantization error on the 1PPS.  It is also normal for 1/T to control the time between phase crossings.  So you have to wait for two miscounts in a row in the same direction to make a change.  And even then, you can't be 100% sure that it's not due to the quantization errors in your 1PPS
signal.

The better GPS receivers will output a quantization error value every second.  But if you're using the 1/T method, there's nothing you can do with it, so you have to live with whatever quantization errors you get.

Anyway, those are my experiences.

Bob - AE6RV

At this point the time measurement is quite crude, with 100-ns resolution.

What you are saying is "With a long enough gate time you can measure
frequency to any desired level of accuracy."

For that to be true the  frequency must remain stable over the long period
you are averaging.  For example let's say you average over 10 seconds.
What happens of the nominal 10MHz oscillator ran at  10.1 Mhz for five
seconds and then 9.9 Mhz for 5 seconds.  You'd measure 10Mhz and be happy.
So there is the problem.  With long period averaging, you can only do it
for so long as you trust the clock to not change.

All the long gate time tells you is the AVERAGE so it only works of the
change over the period is smooth and monotonic.

If you want very good accuracy, say one part in 10^14 you can't get there
by waiting days and weeks because the frequency will move while you are
measuring it.

The quicker way is to measure the phase.  You can still do this digitally.
Make a one bit measurement.  zero if the phase leads, 1 if it lags and
then over time you want the average to be 0.5  This works much faster
because your controller does not have to wait for an entire cycle of error
to accumulate.

Your counting system would have to wait for a one cycle error and then make
a very large and infrequent correction.  It could work for some use cases.

Chris Albertson
Redondo Beach, California

No. The timing resolution per second is always 100 ns. You're probably thinking about average frequency, in which case dividing by T is sometimes valid, and it looks better and better as time goes by, usually.

What saves you here is that your counter noise (100 ns) is likely greater than the quantization noise. So you can pretty much ignore the receiver 1PPS quantization noise. For people with much lower measurement noise (e.g., 1 ns) the quantization noise becomes a more important piece of the error pie.

Try not to say average "out"; that sounds like it goes away over time or gets smaller. You're doing a timing measurement so the 100 ns measurement granularity is always there, on every measurement.

All other factors equal, a GPSDO based on 100 ns measurement resolution can never attain the equivalent of a GPSDO based on 10 ns or 1 ns measurement resolution. Waiting shorter or longer doesn't change the RMS timing accuracy.

/tvb

Tom,

I took his 100ns figure to be simply the period of 10MHz.  He mentioned using an interrupt driven system, so the counts should not necessarily be limited to 100ns accuracy.  At least on the PIC I'm using, the CCP and timer interrupts don't seem to be synchronous with the PIC clock.  I could be mistaken.

Bob

it would be interesting to see the "accuracy" of the 1pps pulses by
comparing them with a second 1pps pulse, which is derived from a
rubidium standard, which on his own does not have quantizing errors,
73
KJ6UHN
Alex

On 2/26/2014 6:03 PM, Bob Stewart wrote: