Controlling maxstep during tstab in Harmonic Balance

I did a bit of checking, and then found an internal discussion about this. With a single tone, maxstep is constrained to be a 25th of the fundamental period - you can't set it smaller. For multi-tone it appears to be (as you said) 100th of the first tone's period - you can set it smaller, but not larger.

Now why would this be? Well, one of the key reasons for the tstab is to generate a good, accurate starting point for the harmonic balance algorithm Realistically you're going to have a fair number of points in the period for that to be useful - so setting it to 100ps (and I appreciate that may have just been an experiment) would only give you fewer than 4 points per period - so that's not going to be anywhere near enough to usefully perform a Fourier analysis to start the harmonic balance from. If you're asking for 10 harmonics of the first tone then at least 20 points would make sense (I should note that the number of harmonics does not affect this maxstep setting during the tstab though).

So put simply, the choice has been made as a practical decision to ensure a good start to the harmonic balance analysis itself. I can understand why you might want to speed up the tstab, but it then becomes significantly less valueable for the conditioning of the hb algorithm. Note that the tstab is only simulating with the first tone enabled, by the way.

A couple of further points:

1. Setting skipdc=yes seems a pretty bad idea to me - usually you'd only do that if either you're starting with an initial conditions file, or have your sources ramping up somehow (hard to do with hb).
2. oversample wouldn't affect the tstab anyway - what that does is add more sample points during the hb interval - it increases the number of points that it converts back into the time domain (via an IFFT), simulates in the time domain at those sampled time points, and then converts back again. This is normally used to increase accuracy when you have switching circuits or square(ish) signals.

Andrew

Thanks for the quick reply, Andrew! And for the pointers - they're helpful too.

I get where you're coming from - why tstab should help condition the hb algorithm correctly. The reasons you suggested make complete sense for cases when tstab is required to stabilize transient effects from enabling Tone-1. 

But, often if not always, tstab is also needed for circuit ramp-up (supply ramp-up due to start-up issues for eg., as in my case). These transients can take several thousand periods of Tone-1 to settle. Since the settling behavior is slow, we can normally increase the transient step sizes comparable to or larger than one Tone-1 period without compromising on accuracy. While running transient analysis, we exploit this by disabling the source for tone-1 (and other tones) during circuit ramp-up and setting maxstep to a corresponding value that is often larger than Tone-1 period. Spectre takes large steps during ramp-up because there are no high-frequency stimuli and the settling behavior of the circuit is rather slow. 

Once the ramp up settles, we enable source of tone-1. Spectre automatically takes smaller steps since tone-1 is high in frequency and the circuit typically also responds faster to Tone-1 (hence the large number of harmonics and/or oversampling ratio for Tone-1). Since Tone-1 is typically large, enabling can often cause another transient that needs to settle before the circuit becomes periodic. However, this transient settles rather quickly since the circuit is typically designed to respond faster to changes in Tone-1 compared to that in supply voltages. So, we only need to allow a small stabilization window before taking measurements on steady state response.

You likely know all this only too well - my apologies if I dragged you through it unnecessarily. Just to share my example, my current simulation is for a wireless receiver that requires 100us+ settling time for supply ramp-up, 1/1000th that for enabling the LO (Tone-1). When using transient analysis, it only takes me a couple minutes to get the circuit to settle before I can take measurements on steady-state behavior. However, when using harmonic balance, at 100 samples a period, spectre is essentially forced to compute close to a million samples during just supply ramp up, needing hours to get through tstab (hb itself only takes a few minutes after that). It will be *really* helpful if we can somehow replicate maxstep control in transient analysis to tstab in harmonic balance analysis as well.

I know little of what that replication really means and I could be easily mistaken about the whole thing. But, I imagine the issue should be a frequent scenario with transceiver or mixed signal (switched capacitor) designs and would be really useful if it can be worked around.

Please let me know if I can provide any further information - will be glad to. Thanks again for responding!

Hi Raghu,

I understand. I think the best thing here would be to contact customer support - sharing this information there (and ideally being able to see the example that is challenging) would be a good idea for us to work with Cadence R&D to figure out the best way of handling this (which almost certainly needs an enhancement). It's far more likely to be done if we can tie this to a real customer with a real example...

Andrew

Hi Raghu,

I understand. I think the best thing here would be to contact customer support - sharing this information there (and ideally being able to see the example that is challenging) would be a good idea for us to work with Cadence R&D to figure out the best way of handling this (which almost certainly needs an enhancement). It's far more likely to be done if we can tie this to a real customer with a real example...

Andrew

Thanks, Andrew. Will do.

Raghu