Spectre X: is "noisefmax" treated differently in the various accuracy modes?

Dear dontpanic,

dontpanic said:

Hi! I am finding that the noise levels in my transient simulations change non-negligibly according to the Spectre X accuracy settings, despite explicitly setting the noisefmax parameter (and noisefmin too).

In a nutshell, it seems that the more accurate the simulation, the more noise is added, as if the tighter time steps somehow would allow for a wider-bandwidth noise injection.

This is a difficult question for me to give specific guidance, but without knowing all the details let me toss out a few comments that may (or may not!0 be relevant dontpanic.

1. I am not sure what you are using as a noise metric to determine the "noise levels". For example, is it based on a time domain waveform observation or a frequency domain based measurement? If a frequency based metric such as an average PSD value or integrated PSD result, have you considered the number of samples chosen for the analysis and their relation to the total simulation time? Figure 1 shows, as an example how the PSD can vary with sample size for resistor noise simulated for 2 us with a noisefmax of 100 GHz.

2. What noise levels are you expecting> Are they on the order of the simulato2 accuracy (established by a number of transient only simulations)?

3. You are correct in that setting noisefmax to a specific value will enforce a maximum time step for the simulator of 0.50/noisefmax. Hence, irrespective of your simulator accuracy setting, the maximum time step shown in the log file should not exceed 0.50/noisefmax.

4. A difficult question to answer that probably needs to be considered is the impact of simulator error (i.e., numerical error) on your noise metric. In other words, simulator settings that have less numerical accuracy may show more or less noise than simulator settings with higher accuracy settings. Simulator numerical error is not solely a function of the integration time step. Further, Cadence has published results comparing Spectre X to Spectre, Spectre APS, and spectre APS++ that show differences in accuracy.  I'm not sure what errpreset you are using in your comparison of spectre aps+ with Spectre X mx. In short, I would select as a  start the most accurate simulator settings for a transient noise simulation to minimize the simulator noise on your result.

Shawn

Figure 1