These are rules of thumbs from circuit board implementation days.  
Basically you don't want to kill the oscillation before it is fully stable.  
See the back ground info below for a better understanding of some of the little known properties of crystals.
Here is what I would do:
If you are designing a oscillator on chip and logic circuits are relatively small,  then I would suggest overriding the control loop with a switch until several clock pulses have been counted then release the control loop override switch.  Or you could override the control loop for a given known period from some other source ie simple relaxation oscillator.  Finally as an alternative you could give the crystal oscillator a good first kick to send the oscillation to steady state faster.


Back ground
Crystals often have a phenomenon called sleeping crystal where there are micro particles attached to the crystal physically.  These micro particles dampen the crystal (like dust on the exterior of a pipe organ) and effectively lower the Q of the crystal prior to first oscillation or after a long period of non oscillation.  Right After oscillation occurs the crystal shakes these particles loose and the Q of the crystal increases.  

J

Crystal kicking does not work reliably and should not be used in a design.

Period.

No conditionals...

and yes, I know there are patents on the concept and people have "tried" to use it in products.... don't go there.

About 6 months back I ended up cleaning up a design for a product where they had attempted to use this. It doesn't work reliably.

Get the crystal Q, and look at the spectral response of that simplified model, trying to "kick" that works relaibly only on a simulator.

I dont have the numbers in front of me at the moment but --- the Q of a crystal is roughly 1E5 or 1E6.

Think of it as a very narrow BW bandpass filter.

To "kick" or stimulate that narrow a  BW requires injecting a signal within a few Hz of accuracy.

Doesn't work in the real world, unless you already have a xtal running at that same frequency, and then, whats the point?

Suggest - look at a normal start crystal (slow rise yes) and provide a gain regulator and output buffer so that you use the output and stabilize the amplitude at a small signal out of the crystal, instead of waiting for te slow amplitude buildup.

The classic single inverter with a crystal around it does not have enough gain in some process corners, unless you push the size of the inverter up.

When you push the inverter size up to get the gain needed in all corners, the static current gets terrible in other corners.

Look at another gain amplifier configuration instead.

I really don't see any good reason to put the bias current of the system in a control loop. Fixed bias conditions should be fine, and you really should not need any other control loops.

It's an amplifier and a crystal, keep it simple.

If you need the faster start time, look into a buffer amplifier (outside the crystal loop) to get a usable signal faster. Depending on your needs you may not need any gain control in the feedback, just let the thing go nonlinear (the buffer gain system) as the crystal signal builds.

Loose electron read this because you may not see the bullets you think you dodged! and reconsider the need to literally kick the dust of the crystal.

stated definition
Low-level nonlinearity is sometimes called
“sleeping sickness” or “second level of drive” Its cause is generally resonator surface condition, which can be controlled through design and
control of manufacturing processes.
found in->
http://www.mtronpti.com/pdf/contentmgmt/crystal_resonator_terminology.pdf

For more info on why
http://www.femto-st.fr/~salzenst/04EFTFbrendel.pdf

I do not understand the reason for having two control loops.

If the reason for the DC bias control loop is to improve startup and increase reliability the only way to derive a control signal is the amplitude themselve or using a counter. But the the counter solution is critical because you have to set the bias start value and final value. That places more risk on variability.

I have done some regulated NMOS and NPN based pierce oscillators in the last 20 years and found that the amplitude regulator need some filtering but the filtering needed for proper operation is surprising low. The remaining peak to peak versus avarage value of the bias current in the regulation loop could reach 30% w/o significant impact on frequency stability and and final amplitude. I design it in the range 3%-10% because there is some phase sensitivity of the bias current modulation to the effective Gm.

To loose-electron:

What the point of tuned kicking XTAL?

There is a big interest in doing that because startup reduction is critical in some radio applications. If the startup time is 100x the active time of a system it contribute significant to the energy budget. In terms of efficiency you want to operate the resonator with specific power level. That requires some energy stored in the reactances. The question is now how to pump the energy in so that the efficiency, reactance energy to DC energy, is high.
If you calculate the effciency you are right that efficiency is invers to the relative frequency difference. But there are kicking oscilators out there which are tuned to ppm and remain there over % of supply and slow temp changes.

[quote author=rf-design link=1266871751/0#11 date=1268126222
What the point of tuned kicking XTAL?
[/quote]

I am very aware of what people want to do with crystal kicking, and I have read some of the papers and patents on the subject. I have also designed enough RF front ends that wanted to be used in a time interleaved fashion and get things started up quickly is something I have had to deal with before.

However, I have also designed enough oscillators to know what works and what does not work.

More specific to crystal kicking, I have had my hands on a design in the lab and seen what happens, what the yield losses are.

They are unreliable, and have high yield losses because you cant "hit the kick frequency" with the needed accuracy.

If you can "crystal kick" with enough frequency accuracy to do something real towards speeding up the oscillation time, that means you already have an oscillator running with a well defined, highly accurate frequency.

If you got that, what need is there in starting up a second oscillator?

Even a tweaked and trimmed RC type oscillator will not be consistent enough to stimulate a crystal properly, thermal variance will make this drift out of band from the crystal.  A tuned LC system will have a little better success but getting things aligned in frequency is a fussy prospect that is not viable in a volume product. Remember, the whole idea here is to create an accurate oscillation, and you don't have one readily available to work with.


I will come back to my original statement - If you want to get an oscillator up and running fast, look into running the device at lower amplitudes to avoid the slow ramp up in amplitude. There are also ways of running crystals at low amplitude, low power setups that can be kept running in the background and consume little power.