Very often I have faced the same problem - and my experience was that it needs at least app. Q cycles until the oscillator comes to a steady state.
I don´t know if there is simulation method to allow a shorter start time.

using PSS just means that the simualtor has to burn the time to find the PSS solution, and that in the background is just running the .tran to find it.

I have inserted a damped sinusoid voltage source in series with the crystal model, which kicks the crystal in the past.

any other ideas?

The initial condition current on the inductor within the crystal model works nicely!

5 star! Thanks!

Visjnoe wrote on Aug 25^th, 2008, 11:07am:

Which envelope analysis do you mean shooting Newton based or Harmonic Balance based ?
If you use RF simulator without any kick start of oscillator, simulation speed is not improved.

If you use shooting newton based, you have to kick start oscillation by 2 or 3 in the above or using step power supply.

If you use HB based without transient asist, you have to set very accurate final state oscillation frequency.

If you use transient asisted HB based, you have to kick start oscillation by 2 or 3 in the above or using step power supply as same as in shooting newton.

Situation is same in both HB based envelope analysis and transient assisted HB analysis.

Anyway you have to invoke any kick start method for very high Q oscillator whatever simulator you use.

Visjnoe wrote on Aug 25^th, 2008, 11:07am:

Using kick start of method 2 or 3, I execute Shooting Newton or HB analysis.
So result is true steady-state.

Visjnoe wrote on Aug 25^th, 2008, 11:07am:

I don't think simulation speed is improved even if you use RF simulator without ank kick start.

If start up time is improved in envelope analysis without any artficial kick start, it is very suspicious.

Maybe you mean compared to common transient analysis, envelope analysis is fast.
If so, speed of envelope analysis is not always superior to transient analysis.

I took a careful look at the envelope methods and did not see a advantadge there. Still need a kick start method.

Since simulations of cold starting crystals take HUGE amounts of simulation time (and memory space into the big GByte world) often the best strategy there is to Initial Condition with a current on the inductor in the crystal to roughly 20-50% of expected final amplitude. Looking at the envelope, if it degenrates, you got gain problems, if its growing you are ok.  Unfortunate but the simulations can take 10-40 hours to run. (and thats on a good UNix server not some old junk PC)

A simple AC analysis on gain is a good guide to whether or not the gain is sufficient, but the transient test (above) needs to be also applied. Gain is heavily bias point dependent, so AC can be misleading. Put some gain margin on your worst case AC gain.

Other little goodies learned on crystal simulations:

AC analysis of a crystal will show a complementary pole/zero response side by side. This can mess the simulation up. For simulation purposes (tranisent and AC) model the crystal with just 3 elements, the series RLC and dont include the parasitic bypass capacitinace. That way you are simulating with just the pole response of the crystal resonance.

The high Q of a crystal can mess up simulations. Drop the Q of the crystal down. Normal Q is 10E4 to 10E6, bring the Q down to 100, to make the simulator happy. That means increasing the resistance in the series RLC.

The ESR of the crystal can play into the loop  gain due to voltage division between the ESR and input impedance of the amplifier.

Due to the above two items, the crystal model gets to be a little more complex than a series RLC circuit. On the input driven side it is a series RLC, buffered by a VCVS (voltage controlled voltage source) which then has the VCVS connected thru a series resistance (the crystal ESR) to the input of your amplifier. A little more complicated but gives viable results.

Starting a crystal (or any other oscillator) is not a function of noise or offsets in simulation. Have a careful read thru section 4.4.2 (page 209) of Ken Kunderts book (The designers guide to Spice and Spectre) for simulation starts and the concept of equilibrium. In the real world (not simulations) the noise/offset plays in, but generally not in simulation.

Crystal oscillators look simple but in reality have lots of touchy issues.

Dear,

I took the time to do some simulation experiments on a 32.768kHz crystal oscillator.

1. The transient analysis (using initial conditions) took 7hrs.

2. The steady-state analysis using harmonic balance took 2min.
    You immediately can assess the steady-state conditions, but unlike transient analysis
     you cannot see the detailed start-up behavior over time.

I can see why the RF simulation using shooting-Newton methods might not gain you that much in simulation time, but it turns out that an harmonic balance simulation speeds things up significantly. This is without decreasing the Q of the crystal oscillator.

Again, I'm not the simulation expert here, I can just say what I empirically see using a commercial RF simulator on a CMOS 32.768kHz oscillator.


Regards

Peter

Visjnoe wrote on Aug 26^th, 2008, 12:21am:

If you have interest only on steady state, you don't have to invoke envelope analysis. Rather just run simple steady state analysis such as HB or Shooting Newton.
Envelope analysis is time varied steady state analysis, so envelope analysis is redundant and slow, compared to simple time invariant steady state analysis.

Visjnoe wrote on Aug 26^th, 2008, 12:21am:

This is dependent on situation for which you apply envelope analysis.

Envelope analysis is efficient if energy is concentrated on one harmonic.
If energies are distributed to many harmonics, envelope analysis is less efficient than common transient analysis.
And if a complex envelope(magnitude and phase) is varied fairly fast, envelope analysis is also less efficient.
These are true for both HB and Shooting Newton based envelope analysis.

Visjnoe wrote on Aug 28^th, 2008, 2:09am:

Your initial conditions are not good.
If you apply proper initial conditions with skipdc=yes(spectre) or uic=yes(hspice), it can reach steady state immediately.
Here initial conditions I mean are all node voltages and all state currents. Not only initial current of inductor.

Visjnoe wrote on Aug 28^th, 2008, 2:09am:

If you have interest on start-up behavior over time, use a time varied steady state analysis such as envelope analysis.

Visjnoe wrote on Aug 28^th, 2008, 2:09am:

Just your initial conditions are not good.

For very high Q osciilator, HB analysis is not always superior than Shooting Newton.

Again you have to consider the followings for rapid simulation of very high Q oscillator.
  HB analysis require very accurate final state oscillation frequency.
  Shooting Newton require good initial conditions.

For transient assisted HB analysis, both good guess of final state oscillation frequency and good initial conditions are useful.