Transient Simulation of Crystal Osc

Dear ART90,

ART90 said:

the problem rises when I tried to study the LDO ripples into the transient simulation using a periodic model file of the ripples in a vpwlf instance as the supply of the circuit. even if this supply is not connected to the crystal and is just present as a floating net. the crystal damps. 

is there any way to properly simulate such a circuit?

vccc1v8_ripple is present in the sch but floating:

I will provide my thoughts, but will need your guidance to establish if the hypothesis is valid.

In looking at your simulator settings, it is clear you have errpreset set to "conservative" - which is great. However, my concern is that when you include the periodic model of the linear supply in your simulation netlist (even without the supply connected to the core oscillator circuit), the resulting simulation integration timesteps are increased relative to the case where you are using an ideal supply to supply the core oscillator circuit. As a result, spectre is not providing an accurate solution.

To verify this, you might compare the "average" timestep by examining the ratio of maximum simulation time/number of steps used in the simulation without the periodic  linear supply model in your netlist to the same ratio of your simulation output log for the netlist without the piecewise linear source

If, however, you only saw this behavior when the periodic source powered the oscillator, then there is a potentially different reason - which I would have detailed without your added information.

As an example, from a recent pss simulation I submitted, from the pss spectre.out file, examine the equivalent  following lines in your spectre.out file

========================================
`pss': time = (100.698 ns -> 100.772 ns)
========================================
pss: time = 100.7 ns (161 m%), step = 119.1 fs (161 m%)

...

Number of accepted pss steps = 3108

Hence, the "average" integration step size is  = (100.772 ns - 100.698 ns)/3108 = 74 ps/3108 = 23.8 fs.

If the two averages are different, I would recommend adding a value of "maxstep" to your simulator options to enforce that sufficient timesteps are included in each 25 MHz oscillator period. The maxstep parameter can be set in the Transient analysis GUI.

Let me know your thoughts ART90.

Shawn

thank u so much, Shawn,

I just tried to be a bit pragmatic! I set maxstep to 4ps in my simulation and that absolutely works. I started with 40ps -> not significant improvements -> the OSC was still damped. but moving down to 4ps -> helps a lot. I would like to do some more investigation, mainly on the average timestep calc. I'll keep posting them here. definitely it's useful to optimize the step size, otherwise, the simulation time would be terribly long! sounds fine for me, as far as it's just a TYP run to do a sanity check on the freq. variation vs. supply ripple. but a full corner run needs a bit of optimization!

/resized-image/__size/1080x0/__key/communityserver-discussions-components-files/33/Screen-Shot-2020_2D00_06_2D00_23-at-03.28.37.png

Thanks again for ur prompt and useful comment,

Alireza,

Dear Ailreza,

I am very glad to hear my thought provided a little insight - thank you for lettng me know! I will add my experience with very high Q oscillators (10,000+) such as you are simulating, I've always needed to make sure there are at least Q/10 or so samples per period of the fundamental frequency. In your case, with a frequency of 24 MHz (41.66 ns), and a Q of 10,000, that suggest a tilmestep of  41.66 ns/10,000 = 4+ ps. This is not a "hard" requirement - but based only on my experiences. Intuitively, in simulations of the phase noise of an oscillator with a Q of 10,000, one is trying to assess phase errors of less -100 dBc/Hz - which is 10^-5 of the period...but this is just, once again, a rule of thumb and just an intuitive basis for my comment.

Shawn