Connect with us

Crystal drift

Discussion in 'Electronic Design' started by Ken Smith, Sep 23, 2004.

Scroll to continue with content
  1. Ken Smith

    Ken Smith Guest

    As I type this, my Stanford Research generator vs. a Digikey 100MHz
    oscillator comparison is drifting by about 5 parts in 10^10 per second.

    While watching this I have a thought. If I could get a hold of the
    temperature set point of an OCXO, and could connect it to a DAC, can I
    make a very slow very low noise sweep generator with a very narrow range
    or did I miss something?

    I need to be able to move a frequency about one part in 10^7 or 10^8 very
    smoothly. Any thoughts?
  2. John Larkin

    John Larkin Guest

    Most any serious OCXO also has a VCO input that you could tap into. Or
    a rubidium, if you're really serious; they generally have an
    electrical tweak input with extreme resolution.

    SRS makes a brick-type SC-cut OCXO with an electrical VCO input; it's
    very stable. $400 roughly, as I recall.

    Or build a quadrature mixer and offset an oscillator output with the
    output of a function generator.

  3. Ken Smith

    Ken Smith Guest

    Oh!. I'll have to check into that. I know that the extreme low noise one
    I had custom made for this project doesn't because the variactor diode
    makes a path for noise to get in. If I can find a fairly low noise one
    with a VCO input that may work for me. I was just now considering a
    non-OCXO and putting in in my own oven for testing.
    Two rubidium clocks with one being tweeked may be what I need. Getting
    the money to buy them may be a bit of an issue. I'll try the $100-$1000
    range solutions first.

    I had rejected that on a noise worry. I'll have to take another look.
  4. John Larkin

    John Larkin Guest

    I got a nice Efratom unit on ebay a while back, $350 or some such. It
    drifts nanoseconds per hour against a cesium, and has an electrical
    tweak input.

    What are you doing?

  5. Ken Smith

    Ken Smith Guest


    I need a good AVAR number for times in the 0.01 to 10 seconds range.
    Can't say yet.
  6. Mike Monett

    Mike Monett Guest

    You probably have a huge library already, but Philips has a brief summary
    on quadrature mixers including the Weaver method at

    One advantage of the Weaver method is a lowpass filter is used instead of a
    phase shift network. This might help when generating small frequency offsets.
    But with any scheme using mixers, the additional spurs and noise may degrade
    the output too much. Please let me know your conclusions - I have a similar
    problem with generating high quality signals with a small frequency offset.

    Rick Karlquist shows a repeated mix-and-divide method using inexpensive
    ceramic filters at

    This may give a low noise output signal, but the method requires one stage for
    each digit in the output, so 10 stages plus shielding might take up some room.

    Rick has updated his web page to include some of his patents pending:


  7. Phil Hobbs

    Phil Hobbs Guest

    Sideband locking is a possible strategy here. If you mix the two oscillators
    together, you can phaselock the beat note to a function generator with a
    frequency-phase detector, e.g. a 4046--it won't lock up to the image
    frequency, because the sign of the loop gain is opposite for the two
    sidebands. By making the loop narrow enough, you can avoid having the beat
    note modulate the VCXO significantly. You do need a large enough frequency
    offset to make this work, and although it won't lock up to the image
    frequency, it can lock to higher IM products if you don't take some care to
    prevent this (e.g. by acquiring lock at a high enough beat frequency that
    only one product is within the tuning range of the VCXO, then sweeping as

    If you need to have the beat frequency go through 0, then John's quadrature
    idea is a good one--otherwise you could mix with a constant offset frequency,
    filter, then do the same sideband locking trick.


    Phil Hobbs
  8. Mike Monett

    Mike Monett Guest

    Phil Hobbs wrote:

    Wouldn't the pll rail if it approached the image from the wrong direction?
    Aren't you locking to the beat frequency? With a phase/frequency detector and a
    balanced charge pump that has zero deadband, there should be very little ripple
    on the output.
    Heh - A frequency offset of 1e-3 or less would probably be quite difficult:)
    Hmm... I'm not sure this came out as clear as you intended. Wouldn't a mixer
    driving a phase/frequency detector ignore the higher IM products, or maybe I'm
    not following you?

    Also, depending on the frequency, a simple D-flop makes an excellent digital
    mixer. Put one input on the clock, the other on the D. The Q output will switch
    at the difference frequency.

  9. Are you stuck to sweeping the OCXO, or could you just sweep a digital frequency synthesizer running
    off the OCXO?
  10. Ken Smith

    Ken Smith Guest

    I need one frequency to be nominally 100MHz the other to be nominally
    about 100KHz.

    The nominal 100KHz needs to have a frequency such that its period is:

    N + 1/M + X cycles of the 100MHz

    N is an integer
    M is a small integer 2,3,4..33
    X is a small positive number on the order of 0.01

    This all has to be done with an AVAR that is about as good as an OCXO will

    I think I have come up with a way to do the measurement more easily. I'm
    going to have to write a little code though.

    I plan to set the SRS generator so that the required conditions are just
    about met. Then I'll leave the experiment running and some software
    watching it. When, due to drift, the frequencies pass through the desired
    values, the software will grab a bunch of data and save it away for me to
    study later. This way, I can take advantage of what is normally a
  11. Ken Smith

    Ken Smith Guest

    I am using a synthesizer but I need about 12 digits of adjustment. By
    picking the right numbers I can make the frequency something like:

    123456.7000008 KHz

    so nudging the OCXO can give me the bottom digit and the synthesizer can
    give me the upper ones.
  12. Phil Hobbs

    Phil Hobbs Guest

    It won't rail, but it will try to lock up at the wrong null--which in a PFD
    is a huge cliff due to the sawtooth characteristic, leading to a much larger
    loop bandwidth in this condition. If this is done right, the loop will not
    be stable at this increased gain, and so it will get kicked away. Whether it
    gets kicked in the right direction eventually is a design issue--I wouldn't
    build something like this without aided acquisition. My favourite
    acquisition aid is an auxiliary positive-feedback network around the (active)
    loop filter--when the loop is out of lock, it oscillates slowly until lock is
    found, at which point the negative feedback overwhelms the positive, and lock
    is acquired. All it takes is a twin-T or a phase shift oscillator network.
    Yes, but not *none*. This is an ultraprecise application, after all.
    Well, 1e-3 fractional change in 10 MHz is 10 kHz, which is way more than
    enough. If you mean 1e-3 Hz, then I agree--a second oscillator to provide the
    offset would be needed, as below.
    It depends on what the filtering following the mixer looks like. If the main
    beat note is outside the filter bandwidth, it's quite possible for a
    higher-order term to cause locking with a PFD, especially if there's a
    limiter after the lowpass filter.
    Not with a PFD--metastability will blow you right out of the water. Every
    lost cycle equals lost lock. The D flipflop trick can work with a narrow loop
    and a multiplying phase detector though--I used it when I built a pilot tone
    generator for what I think was the first commercial direct broadcast satcom
    system, in about 1982. (I had just got my bachelor's degree in astronomy and
    physics at the time, and they hired me to look after all the ultrastable PLL
    stuff--talk about being thrown in the deep end. For the frequency reference
    board (different from the PTG) I had to invent a fractional-N synthesizer
    based on resynchronized rate multipliers. It worked great, eventually.)


    Phil Hobbs
  13. Well, since it's the area I work in (if you only have a hammer, every problem looks like a nail), I
    was thinking of a freerunning telecom DPLL. Just use it as a frequency synthesizer, usually the
    frequency offset or holdover value can be programmed to a very fine degree. Use the OCXO (possibly
    divided) to drive the system, then you can set a very fine offset between the two. But if you need
    to trim the 100MHz clock as well, you'd need some other way. Is it just a differential frequency
    difference that you need?
  14. John Larkin

    John Larkin Guest

    This one

    uses an EclipsLite ff as the phase detector, locking the local 155 MHz
    rock to the incoming OC-3 data stream. BW is about 10 KHz for acquire,
    2 KHz when tracking, with the uP switching loop filters when it feels
    it ought to. About 200 of these are scattered all over the NIF laser
    to fire everything at the right time. You can barely make out the
    crystal oscillator sitting up on tiny springs.

    Biggest hassle was the (expletive deleted) 850 nm VCSELs for the
    optical-out option. What nasty parts!

  15. Ken Smith

    Ken Smith Guest

    I can nudge either input for the fine adjustment and the circuit won't
    know which I've moved. I can't move the 100MHz by 0.1% or anything like
    thant though.
  16. Phil Hobbs

    Phil Hobbs Guest

    Interesting. Is this just a single bang-bang detector, or something like an
    early/late gate? If it's a bang-bang loop, how do you handle metastability?
    The loop is going to be trying desperately hard to create perfect
    conditions for it. Of course, it's not as bad a problem using the FF as the
    phase detector as when driving a PFD from a FF mixer.


  17. John Larkin

    John Larkin Guest

    I'm using a single d-flop, with the incoming (biphase) optical data
    clocking it and the local crystal clock as the D input. So every data
    edge makes an early/late decision for me. The ff output is just
    lowpassed (p+i) into the VCXO. I like this because I'm after extreme
    time stability, equivalent to a tiny fraction of a degree, and this
    does it.
    Well, I look at it this way: the FF output is going to go either high
    or low every shot, and stay that way till the next data edge. If it
    goes metastable for a couple of ns, that means the timing was right on
    the early/late edge, so it didn't matter if it came out 1 or 0, so who
    cares if it splits the difference? An EL51 is going to resolve pretty
    quick anyhow, and one sampling cycle will have a miniscule influence
    on a 2KHz lowpass filter and a narrow-range VCXO.

    Hey, I'm an engineer. Whatever works.

  18. Mike Monett

    Mike Monett Guest

    Sure it will. Consider a lower sideband system with a reference
    signal at 10 MHz, a vco at 9,999,999 Hz, and a pfd at 1 Hz.

    Case 1: If the vco frequency decreases, the difference frequency
    between the reference and the vco increases. Therefore the pll
    must increase the vco frequency to regain lock.

    Case 2: If the vco frequency goes above the 1 Hz target, the
    difference will decrease, and the pll must decrease the vco
    frequency to regain lock.

    If the vco frequency is well above the reference, say at 10,001,000
    Hz, the difference is greater than the 1 Hz target. This will cause
    the vco frequency to increase as in Case 1. However, it is already
    too high and the loop will rail.

    If the vco is above the reference, but below the 1 Hz target, the
    difference is below 1 Hz and the pll will decrease the vco frequency
    as in Case 2. The mixer output goes to zero when the vco is equal to
    the reference, and the pll output will decrease the vco frequency as
    before. The vco will then cross the reference and lock on the proper

    So it is important to approach lock from the correct direction.
    We are talking about very low offset frequencies. Any reference
    spurs on the vco will be buried in the close-in phase noise.

    Definitely not true. A missing or extra transition will have little
    or no effect in a narrowband loop since it won't have time to move
    far. The clocks will remain within the feedback reset time and the
    problem will get corrected on the next cycle.

    The biggest problem with the digital mixer is multiple transitions
    at low offset frequencies due to phase jitter.

    For example, the time delta between samples is

    Delta = ------


    Offset = Offset Frequency in Hz
    Fs = Signal Frequency in Hz

    A 1 Hz offset at 10 MHz gives a time delta of 1 / 1e14 = 10 fs. So
    every 100 ns, the sampling point moves over 10 fs. Since most
    signals have rms jitter values much higher than this, the digital
    mixer will produce a string of random transitions as the phase goes
    through zero. This will mess up a pfd or any trigger circuit. The
    solution is to fire a single-shot on the first transition and ignore
    the rest.

    The Binary Sampler test circuit uses a single shot on the output of
    the digital mixer in Fig. 1, but it is not shown due to lack of
    Mike Monett
Ask a Question
Want to reply to this thread or ask your own question?
You'll need to choose a username for the site, which only take a couple of moments (here). After that, you can post your question and our members will help you out.
Electronics Point Logo
Continue to site
Quote of the day