🚀 go-pugleaf

In article <571@cpocd2.UUCP> howard@cpocd2.UUCP (Howard A. Landman) writes:
>(3) There are several commercial companies (Meta-software with HSPICE, Shiva
>    with Power-SPICE, ...) who would be happy to charge you around $50K for
>    a slightly enhanced version of SPICE which they claim will always converge.

Both these products do converge much better than vanilla SPICE 2G6 and neither
claims their product will "always converge."  My estimate is that they will
converge at least 5-10 times better for MOS circuits -- meaning that no more
than one-fifth to one-tenth as many circuits won't converge.
PowerSPICE has a very good convergence ability without much parameter
tweaking.  The best strategy for a potential buyer is to submit very
difficult benchmark circuits to these companies.  Good benchmarks are
circuits that reflect what you are designing and that SPICE can't
converge on, or gets obviously wrong answers for.  See how these two
products fare in terms of accuracy and speed.  If you have chosen your
benchmarks well, you shouldn't have to "pray" much.

By the way, PowerSPICE is not a "slightly enhanced version of SPICE" --
it is written in C completely from scratch with entirely different, more
reliable algorithms and improved models than SPICE 2G6.  It is much faster
than SPICE 2G6 on virtually all MOS circuits; but then MetaSoftware works with
you to develop custom models for the fab process you will be using for your
device.

Convergence problems come from a host of sources, most commonly discontinuities
in the device models and poorly chosen parameters for the integration
algorithms.  There are tradeoffs here --- device models that are guaranteed
to be sufficiently continuous are almost always much too expensive to
evaluate, or have other problems such as floating point over/underflow in
regions where they aren't intended to be valid.  Another major tradeoff
is that you'd like to take big aggressive time steps in your integration
algorithm to cut down execution time, but this opens up the possibility
of converging to incorrect values.  In general, good convergence in a
circuit simulator requires:

	1. Bug-free algorithm code (incl. matrix solution package)
	2. Thoroughly tested and characterized device models
	3. A sound understanding and characterization of the algorithms and
		the tradeoffs made to gain execution speed

You can see why vanilla SPICE 2G6 isn't too great at convergence -- there
never really were many people on the project (typical of research!) and
there is no reason for the UC Berkeley research group to find the last little
problems.  It's much more worthwhile for them to go on to new ideas.

Caveat: I worked on the development of PowerSPICE at Shiva (now known as
SimuCAD), and am now a user of HSPICE because of its custom models.  I am
attempting to provide as unbiased a view as possible.  In any case, I defer
any further inquiries to the sales/support staff of both MetaSoftware and
SimuCAD.
--
	Mike Klein		klein@sun.{arpa,com}
	Sun Microsystems, Inc.	{ucbvax,hplabs,ihnp4,seismo}!sun!klein
	Mountain View, CA