Re: [ng-spice-devel] Re: snapshot/resimulate


To ng-spice-devel@ieee.ing.uniroma1.it
From Ed Hudson <elh@spnet.com>
Date Wed, 04 Apr 2001 09:48:01 -0700
cc elh@spnet.com
Delivered-To mailing list ng-spice-devel@ieee.ing.uniroma1.it
In-Reply-To Your message of "Wed, 04 Apr 2001 17:12:59 BST." <CF654D964573D311A1CA0090278A36FF4CBDFC@EDIN_EXM1 >
Mailing-List contact ng-spice-devel-help@ieee.ing.uniroma1.it; run by ezmlm
Reply-To ng-spice-devel@ieee.ing.uniroma1.it


>  but Steve's right that each timepoint
>  would need to be stored as two (or more) complete snapshots,
>  to get both the voltage/current state and the dV/dI state, in
>  order to let spice carry on accurately from where it left off.


        if each snapshot took the equivalent of:

                100TimePoints * (NumNodes + NumBranches)

        then if the simulator only saves one out of a thousand
        TimePoints, or one out of 10thousand, there is a 5
        or 50x increase in the size of the 100% accurate
        simulation results file attainable (assuming
        NumBranches ~= NumNodes)

        this savings factor could be dialed in by the user
        via a .options statement, or perhaps controlled more
        finely (snapshot mode to t=50ms, per-time-point-mode
        next 1us, snapshow-mode after, etc).

        if it has taken me one day to simulate to the point of
        interest, it is ok if it takes 5minutes (ie, normal simulation
        speed) to resimulate the few nanoseconds i may be interested
        in 'out-there'.

        and yes, a snapshot restartable viewpoint, with controllable
        accuracy, would allow a user to run the viewer in the mode:
        step out to where i think the problem is, look around
        quickly to validate you're in the right place, zoom in,
        now resimulate/display more accurately...

> >       for example, might have 100 active nodes and a million 
> > stable nodes...,
> > ...
> >       after initial convergence).
> Unfortunately, spice recalculates the whole circuit even if
> only one node is changing. It helps a bit if most of the circuit
> is stable, but not much.

        well, in my experience, the number of iterations necessary
        at each time point for time-point convergence is a function
        of the circuit activity at each time point, and this effects
        the total cpu time required at each time point.  otherwise,
        each time point in a transient analysis would cost as much
        cpu time as dc convergence.

        for transient analysis, commercial simulators like hspice
        speed up significantly when a circuit is stable, and slow
        down again when activity starts up again.  spice3 has this
        behaviour as well, and infact, when it simulates something
        correctly, usually is faster than hspice...

                -elh

Partial thread listing: