Waveforms and Cables for Precision Timing Measurements

[rubidium]

Throughout the past year I've been bitten by the precision timing/frequency bug. I'm a relative newbie at this so bear with me, but at least I have acquired a variety of stable sources (Cs, Rb, OCXOs) and more recently some signal analyzers that perform far better than my venerable Agilent 53132A - including the Wavecrest DTS-2070 and, more recently, the Wavecrest SIA-3600C. The latter 2 instruments are so sensitive at making time interval measurements between 2 channels that they have opened up a whole new set of questions for me: the quality of coaxial cables and fittings and the nature of waveforms used when making precision timing measurements. Stable sources are 1 thing - actually measuring how stable they are is a whole other world that I am just entering.

As a baseline for addressing these concerns, I simply tried using TimeLab to make ADEV measurements of the noise floors of various equipment setups. With my 53132A used as a TIC, the best I could do was about 4.6E-10 at 1 sec. Then I graduated to the DTS-2070 and got that down to 4.5E-12 @ 1 sec. Lastly, the best I could do with the SIA-3600C was 3.8E-13 @ 1 sec. Playing around further with the SIA-3600C is what opened up my eyes regarding signal properties. (I can measure the increase in delay between 2 channels viewing the same signal by just loosening 1 SMA connector by as little as 1/8-th of a turn! That's how good this system is!) The lowest noise floor measurements that I got with is were with the internal 10 MHz square wave that is available as a "calibration signal." It has rise and fall times in the 200-250 ps range. When I repeated noise floor measurements with a 10 MHz square wave from an external Agilent 32120A functional generator I got results that were far worse. The rise/fall times were about 1.5 ns in that case. Still worse results were obtained when switching to a 10 MHz sine wave. I understand all this, as timing events in these instruments are marked by the crossing of a (generally user-defined) voltage threshold by a waveform's edge, and if there is any slope to that edge the moments of threshold crossings are corrupted by noise. But what I don't understand is how people actually deal with this! So many - even very high-end - oscillators that are sought after by precision timing aficionados have 10 MHz sinusoidal outputs. Does it make sense to convert these to fast rise/fall time square waves? If so, how. An ultrafast comparator, like the ADCMP580 series? Doesn't that just make for an equivalent threshold-crossing problem in the presence of noise? Even then, coax cable dispersion will take even a textbook square wave and corrupt the rise/fall times. So which cables/connectors should one use?

Cheers,
Jim

[thermistor-guy]

...So many - even very high-end - oscillators that are sought after by precision timing aficionados have 10 MHz sinusoidal outputs. Does it make sense to convert these to fast rise/fall time square waves? If so, how. An ultrafast comparator, like the ADCMP580 series? Doesn't that just make for an equivalent threshold-crossing problem in the presence of noise? ...

The Design of Low Jitter Hard Limiters, Oliver Collins, IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 44, NO. 5, MAY 1996
https://ieeexplore.ieee.org/document/494304

A google search for the paper also gives this link:
http://www.ko4bb.com/getsimple/index.php?id=bruces-zero-crossing-detectors

[5065AGuru]

Jim,

Since you are so obviously bitten by the time/frequency bug you might want to play with a DMTD system.
My best one has around 6X10-14th at 1 Sec, and my old favorite an NBS 106B reaches 9X10-14th with care.
They easily take measurements between two stable 5 or 10Mhz sinewave oscillators. read my DMTD tutorial post and also the DMTD board post.
Also but a bit pricey are Microsemi's 3120A and the older TimPod, I think there is another newer one out also but can't find the reference to it.

Have Fun!

Corby

[rubidium]

Got my eye on the thread Corby! 

[edpalmer42]

The first thing you have to do is get rid of sine waves.  You're correct that noise on the signal will corrupt the readings.  But at these levels, sine to square conversion is a black art that rather quickly reaches it's limit.  You can use whatever type of high-speed witchcraft you want, it's just not going to get any better.  I use a simple 74AC04 squaring / 74HC4059 dividing circuit.  I have a Wavecrest DTS-2077 but, like you, I haven't been able to make external ADEV measurements better than about 5e-12@1sec.  A better squaring circuit would improve that, but not by much.  I can't even tell you if the limit is due to my squaring circuit or the DTS.  Maybe the squaring circuit is fine and the DTS is the limiting factor.  It's just too hard to make direct measurements at those levels.  And the DTS-2077 is so LOUD that I just don't like using it at all!

The only way to get better measurements is to change to a different technology.  That's why things like single and dual mixer circuits are used.  They take advantage of the resolution multiplication that results when you mix down the signals.  As you can see from Corby's circuit, it's not hard to get an order of magnitude improvement over a simple squaring circuit. 

Traditionally, in a DMTD system, the magic happens in the ZCD (Zero Crossing Detector).  So if you're optimizing that function, you refer to Collin's paper.  Corby's circuit is interesting because of the input circuit that drives the mixer with a square wave.  I searched the net and found lots of discussion on driving the LO input with a square wave, but nothing on driving the RF input.  It will be interesting to experiment with different front-ends for Corby's circuit.  There are many topologies for sine-square converters that could be dropped in.

An interesting question is when will we reach the limit of ZCD circuits and what comes next?  I don't know what kind of technology is being used in today's labs so maybe we've already hit that limit.

Ed

[jpb]

I hesitate to put this as it is still very much at the experimental stage, but the best results in term of low noise floor I've got using two phase-locked oscillators at 2GHz and 2020MHz and mixing the result to give 20MHz so it is 10MHz in and 20MHz out but with a multiplying factor of 100 (relative to the 20MHz) and then use my FCA3100 counter.

I got results for my two GPSDOs that were what I expected (agreed with measurements using other methods) but when I tried measuring two Rubidiums against each other there is an odd lump in the curves (quite repeatable) so I am uncertain if it is my Rubidiums or my measurement setup.

I realise dual mixer would be better but I don't have two 2 GHz oscillators or two 2020MHz ones.

As to cables I use semi-rigid which is supposed to be best for measurements where very small time differences matter.

[edpalmer42]

I'd need more info to make an intelligent guess, but as a dumb guess, a single lump in an ADEV curve for a GPSDO often means that the PLL isn't tuned properly for that system.  e.g. http://www.leapsecond.com/pages/tbolt-tc/

Are your phase-locked oscillators identical?  Except for the output frequencies, are the output spectrums the same?

I can see that semi-rigid or hardlines would be helpful at GHz frequencies, but I don't know if they'd be necessary at 10 MHz.  But you do have to make sure that the temperature of the cables stays constant and they can't move around so maybe semi-rigid would be helpful, although difficult to accomplish in most cases.  I certainly consider double-shielded, or better, cables to be required.

[tkamiya]

One thing to be aware is that to pass a square wave exactly, theoretically, you'll need an infinite bandwidth.  In theory, square wave is a series of odd integer harmonics.  Otherwise, rise time won't be as vertical, zero crossing won't be as instantaneous, maxima and minima won't be as flat, and corner won't be as sharp.  Realistically, you'll just need a good set of coax and good everything in its path.  On the other hand, sine wave will only need to pass the fundamental frequency.

This is something engineers argued over lunch for decades.  As far as I know, there is no definite answer.  It all "depends".  In my own setup, which is quite limited, I try to use good coax, good connectors, and try to minimize random fluctuations.  Where I can manage it, I use semi-rigid.  For example, I just built Corby's DMTD system.  I used semi-rigid for input side.  Of course, I will use good flexible coax for external connections, one can argue it wasn't necessary.  But it made sense to me to eliminate possible "contamination" of data. 

I took apart HP5071A with a bad tube.  It had a thick semi-rigid coax connection from OCXO to a doubler module.  As it cost a lot of money to do this, I don't think HP/Agilent did this without a reason.  (it is a sine wave there, by the way)  If you look at most high-end instruments, they do use semi-rigid hardware quite a lot. 

[jpb]

I'd need more info to make an intelligent guess, but as a dumb guess, a single lump in an ADEV curve for a GPSDO often means that the PLL isn't tuned properly for that system.  e.g. http://www.leapsecond.com/pages/tbolt-tc/

Are your phase-locked oscillators identical?  Except for the output frequencies, are the output spectrums the same?

I can see that semi-rigid or hardlines would be helpful at GHz frequencies, but I don't know if they'd be necessary at 10 MHz.  But you do have to make sure that the temperature of the cables stays constant and they can't move around so maybe semi-rigid would be helpful, although difficult to accomplish in most cases.  I certainly consider double-shielded, or better, cables to be required.

Thanks for the response.
The phase locked oscillators are fixed commercial ones (a CTI one and a MITEQ one) so they are not identical and I've no control over the PLL settings/filter but when I measure a 10MHz source against itself I get a smoothish and sensible  noise curve (or straight line rather than curve). Also when I measure one Star4+ GPSDO against another the curve I get is identical to that I get with more direct measurements at higher tau and I get agreement down to about 1 second with different measurements so it all seemed to be working well. It is only when I measure one Rubidium LPRO against another that I get an odd lump around 10 secs I think it was (going from memory). I suspect it might be the Rubidiums behaving oddly.
The link you sent seems to be more about the timing of GPSDOs control loop and I get similar results with varying the GPSDO timing on the Star 4s - perhaps the Rubidiums have a similar issue with locking the 20MHz crystal to the physics unit? (I'm on doubtful ground here - as it has been a while since I read up on how they work) so it may well be a genuine measurement of the Rubidium's behaviour except for the fact that published curves I've seen don't show such behaviour.

[FriedLogic]

The first thing you have to do is get rid of sine waves.

   Unfortunately that also tends to create a lot of wideband noise. Having the equipment close, and using short good cables and connectors and quiet power supplies would be a good start. If you happen to be in a noisy RF environment you might need shielding too.
   Another complication is that some oscillators are quite sensitive to injected noise, and square waves and switching transients from the drivers are good noise sources.
   It can be useful to convert to a square wave sometimes, but it's worth thinking about potential issues too.

[edpalmer42]

Regardless of the fact that you can't get a perfect square wave or that square waves create noise, the fact is that square waves give better results when measuring ADEV.  The attached graph shows noise floors for different inputs.  The 'dead Racal' is a 1992 counter with a dead custom chip.  It turns out that external references are squared up with an MC10116 ECL line receiver with a 2 ns rise & fall time.  All other signals are sine waves.

As you can see, the square wave gives better results and, even with the sine waves, higher frequencies (i.e. faster 'rise' times) gives better results.  The two 'dead Racal' graphs are actually on top of each other.  This demonstrates that squaring the inputs removes the penalty imposed on lower frequencies - a welcome bonus!

But even the 2 ns rise time isn't enough to reach the potential of the DTS-2077.  The 1 GHz signal was the closest I could come to a **really** fast rise time.  Given a 1 ns. cycle time, and a very loose definition of rise time, you can think of it as simulating a 400 ps rise time.  Now you're approaching the limits of the Wavecrest's performance.

Note that this only applies to a simple TIC measurement.  It does NOT apply to a DMTD or similar measurement configuration.

[edpalmer42]

I'd need more info to make an intelligent guess, but as a dumb guess, a single lump in an ADEV curve for a GPSDO often means that the PLL isn't tuned properly for that system.  e.g. http://www.leapsecond.com/pages/tbolt-tc/

Are your phase-locked oscillators identical?  Except for the output frequencies, are the output spectrums the same?

I can see that semi-rigid or hardlines would be helpful at GHz frequencies, but I don't know if they'd be necessary at 10 MHz.  But you do have to make sure that the temperature of the cables stays constant and they can't move around so maybe semi-rigid would be helpful, although difficult to accomplish in most cases.  I certainly consider double-shielded, or better, cables to be required.

Thanks for the response.
The phase locked oscillators are fixed commercial ones (a CTI one and a MITEQ one) so they are not identical and I've no control over the PLL settings/filter but when I measure a 10MHz source against itself I get a smoothish and sensible  noise curve (or straight line rather than curve). Also when I measure one Star4+ GPSDO against another the curve I get is identical to that I get with more direct measurements at higher tau and I get agreement down to about 1 second with different measurements so it all seemed to be working well. It is only when I measure one Rubidium LPRO against another that I get an odd lump around 10 secs I think it was (going from memory). I suspect it might be the Rubidiums behaving oddly.
The link you sent seems to be more about the timing of GPSDOs control loop and I get similar results with varying the GPSDO timing on the Star 4s - perhaps the Rubidiums have a similar issue with locking the 20MHz crystal to the physics unit? (I'm on doubtful ground here - as it has been a while since I read up on how they work) so it may well be a genuine measurement of the Rubidium's behaviour except for the fact that published curves I've seen don't show such behaviour.

Remember - I did say it was a dumb guess! 

It's possible that your multiplication is allowing you to see something that's normally invisible.

I don't know if it's possible for two PLLs in series to interact with each other in bad ways.

Unfortunately, it's a case of too little knowledge interacting with too little data - a very bad combination.

[rubidium]

I thought it would be interesting to prepare a family of plots similar to Ed's and compare the Wavecrest DTS-207x series with the SIA-3x00 series - with regard to sensitivity to different waveform edge slew rates. This plot was made with the SIA-3600C (which supposedly has 6GHz front-end bandwidth). What I did was set up a test for the noise floor by connecting a test signal to both input channels via a coax T at the first channel. The 3600C was configured to measure the delay between rising edge zero crossings of the 2 channels). The test signals were 10MHz, 100MHz, 1GHz and 3GHz sine waves, all at 8.4dBm, from an HP-8664A with the low-noise option. The 8664A was connected to a PRS-10 as external reference. The 3600C provides several calibration signals, among which is a nice 10MHz square wave having 150ps rise and fall times (20%-80%). (The only instrument I have capable of actually measuring that rise time is the 3600C itself, which is how I got the value.)

Of course, the results clearly show that lower ADEV can be achieved with faster waveform edge slew rate, just like Ed's plots demonstrate. Notice that the "real" noise floor of the instrument itself becomes apparent in the lower 2 traces. There's no difference between the 3GHz sine (to which we would ordinarily associate a 20-80 rise time of ~0.3/period = 333ps) and the 10MHz square with 150ps rise time. I would say that there's no more to be had with this instrument, due to internal noise, input bandwidth, and probably dispersion in my cables).

The 3600C is outperforming Ed's 2077, except for the case of the 10MHz sine, where the 2 devices are comparable or the 3600C even a couple of dB worse. The latter is something I can't explain, other than the 3600C "doesn't like" lower slew rates as far as timing measurements are concerned. The other interesting observation is with the 2 lower traces, for which there appears to be a mild departure from linearity in the ADEV log-log curves below about tau=2 sec. Can't explain that either.

These Wavecrest instruments are really amazing devices - particularly when you consider their vintage. (I suppose they originally cost as much as a small house back in the day. They still make the SIA-3600's and I saw a brochure that they can be upgraded to the 4000 series for a starting price of $36495!) Despite all of this, it's quite clear that this sensitivity to input slew rate becomes rather moot with the DMTD approach to measurement. That's the way to go.

Jim

[edpalmer42]

So then, all you need is a squaring circuit with a rise time of <= 200 ps and you won't need a DMTD.  Piece of cake!   

One thing I'm paranoid about when making measurements at these levels is impedance matching.  The last thing I want is reflections messing up the signals.  For tests like these, I never use a BNC T connector.  I always use a 50 ohm splitter and different lengths of cable.  This guarantees both a good impedance match and a time difference between the waveforms.

[rubidium]

So then, all you need is a squaring circuit with a rise time of <= 200 ps and you won't need a DMTD.  Piece of cake!   

One thing I'm paranoid about when making measurements at these levels is impedance matching.  The last thing I want is reflections messing up the signals.  For tests like these, I never use a BNC T connector.  I always use a 50 ohm splitter and different lengths of cable.  This guarantees both a good impedance match and a time difference between the waveforms.

Good point about a proper splitter. I didn't have one on hand at the time and so all I could do was use a T:

            Ch-1                      Ch-2
               ^                             ^
sig------->|<-----------------------|
             "T"

There was about 6 ns of delay between the channels.

I just did this test to see what I could get with this system. I remain an advicate of the DMTD to mitigate most of the sensitivities with these measurements to waveforms, and plan to build one once I learn more.
Jim