Anna wrote on Sep 5^th, 2007, 12:58pm:


Hi Anna,

If I understand correctly, you are using a DFT to find the frequency of your oscillator. If you want an accuracy of 1 ppm in estimating this, you would need an enormous number of points to guarantee that your FFT bin width is smaller than the target resolution. Also, it would be necessary to guarantee that you take exactly an integer number of cycles (-1 FFT point) while doing FFT because you cannot use windowing if the accuracy must be that high. But to do this, you would need to know the frequency a priori : ) However, if you are measuring the frequency with an accuracy of 15 ppm, then you probably are already following most such guidelines to begin with.

If you are doing a DFT on the time domain simulation output, I would suggest instead to try and find the frequency directly from this output (also non trivial).

One way to simulate this correctly would be to make the simulator timestep < 1ppm of the timeperiod but this will slow down the simulation. I would also set the integration method to "traponly". However, I do not know an easy method for estimating the frequency to such a high level of precision.

Please let us know if you find a good method.

Regards
Vivek

Crystal oscillators, being very high Q, take a very long time to settle. So with a normal transient, you may be looking at the results while still being far from steady-state. As such, I would tend to trust the PSS results more than the transient results.

You can measure the period of your signal very accurately using the "period measurement" module that you find on the verilog-ams page (http://www.designers-guide.org/VerilogAMS/).

-Ken

Thank all of you for your help.
I measured the period by applying long transeint as intial condition for shorter second transeint which has maxstep=.1ps( which is 1ppm).

I also have seen differences between TR and PSS for high Q circuits. It seems that for tough convergence setting for PSS the initial transient  time approaches the initial settling time of the normal transient simulation.

There is always a difference between AC and TR frequency because of the frequency shift effects of the nonlinear caps in the circuit. TR and PSS should give the right value. This discrepancy is also a good indication that your oscillation frequency is sensitive to amplitude variations.

To extract 1ppm frequency you do not need 1 million cycles. If the set the FFT you can extract the frequency of the oscillation by bilinear interpolation between two neighborhood FFT bins taking the simply the amplitude. The interpolation is accurate up to about the harmonic level. So with -30dB to can extract up to 3% bin size.