fran2k5,

You are right about the window size and the overlapping segments. But the overlapping segments are selected from the total time record. The total time record consists of uniformly sampled time points and the total number of points is specified by the "Number of Samples" parameter. Spectre does not generate uniform time steps by default. However, the Fourier analysis in the psd function requires uniform samples. The Number of Samples determines how many uniformly sampled points to interpolate from the complete (non-uniformly sampled) record for use with the overlapping segments. The window width determines the frequency resolution. The Number of Samples, together with the start and stop times, determines the sampling period, which in turn determines the maximum frequency of the psd analysis.


-Jess

fran2k5,

Cadence displays the psd in units of Volts**2/Hz for the vertical axis and Hz for the horizontal axis. Consequently, you can mix and match psd parameters and the overlaid plots still make for fair comparisons.

One thing that is not obvious is that the psd displays a double sided density; if you integrate only over positive frequencies to compute the rms value, you will need to double the integral.

-Jess

fran2k5,

One trick I use when studying noise floors is to strobe the data. Set the strobe period under transient options to be consistent with the number of points you plan to use in the psd so that the psd function does not have to interpolate. This will drop the noise floor considerably and perhaps make your comparison more robust. As it was explained to me, in different words, strobing sweeps numerical noise under the "signal rug"; the noise still exists but it is in the signal lines instead of the noise floor. As part of the signal, the noise should have less of an effect on the SNR and thus give you a more robust comparison...I think. I suspect it would be even better if you could find a strobe period that lines up with both ADC rates.

-Jess

Thank you very much to both of you!!
I finally understood the right way of proceeding.

I need to perform the comparison keeping the same bin NBW for the 2 converters. The bin size can be kept the same if the simulation time is the same and if the ratio between Nunber of FFT points and window size is kept constant. For the same time, in the converter that goes faster I need more points because of the higher sampling rate. If I consider a number of FFT points that satisfy both cases (equivalent to a sampling frequency for the FFT higher than the highest ADC sampling frequency) it will mean considering more interpolated points for the slower converter and therefore more noise. As Jess suggested, strobing at the actual frequency of the converter gets rid of this problem:for the same simulation time the faster ADC will give more points and as consequence if I keep the n.ro points/window size the same in both cases I will get the same bin NBW.

Thank you again,

fran2k5