dwatola@NEXTASY2.EECS.WSU.EDU (David Watola) (12/02/90)
someone writes:
In addition, I have had some very good results with CAzM from MCNC
(contact jeri@mcnc.org, $300 university price for simulator and "sigview"
X11 graphics). CAzM is a table-based version of spice. I confirmed that
it runs over ten times faster for my MOSIS tiny-chip simulations compared
to Spice 3c1, and it has always converged! A special "powerup" feature
eliminates DC convergence problems, and I have not experienced any
"timestep too small" problems common to spice3c1 derivative glitches.
With CAzM it is practical for me to "spice" an entire 2mm tiny-chip for
submission to MOSIS. I simulated a complete 9,000 transistor chip
on a Sun Sparcstation. The job required about 10MB and computed at a
rate of 10-20 nanoseconds per hour of CPU time.
Incremental results are available as the simulation progresses.
Output can be generated for sigview or for nutmeg.
I checked the
accuracy of smaller simulations against spice3c1, and came up with
identical results.
I am considering dispensing entirely with logic-level
simulators, and going to entire-chip simulations with CAzM.
i also have nothing but praise to give to cazm. all of my digital tiny-chips
will be simulated with cazm now, especially since i discovered some nasty
nasty bugs in irsim. the new version (4.3) is even better than the last.
however, cazm is not a panacea. it does have some limitations. for example,
to use the table-driven mode of simulation, the device curves must be
monotonic--this has presented major problems for simulating scmos with
mosis 1.2 micron parameters (you can work around this by using the model
equation evaluation mode rather than table mode, but then you basically
have the speed of spice). i HAVE had the problem of 'timestep too small'
with cazm, but only on analog circuits--the fix is the same as in spice;
change some of the option parameters at the expense of simulation speed.
finally, one MAJOR problem: i am using cazm to develop models for
floating gate fets fabricated thru mosis' double polysilicon process.
it turns out that, because of the way cazm updates node currents and
charges, it is possible for even a simple model (say a nonlinear resistor)
to lost charge or current conservation--i.e. you can get a finite current
in one end of the resistor but nothing out the other--even with 'reasonable'
options. changing charge tolerances, etc., fixes this but there is no
warning from cazm. of course, if you know how your simulation results
should end up looking, it is easy to track it down.
once again, these problems are nasty, but easily avoided. cazm is an
excellent tool, and i highly recommend for both analog and digital circuit
simulation.
dave watola
dwatola@nextasy2.eecs.wsu.edubedk@ICSI.Berkeley.EDU (Brian E. D. Kingsbury) (12/14/90)
My sincerest apologies for bringing this up so soon after it had been discussed here... Could some kind soul out there please send me information on getting a copy of the circuit simulator CAzM, and (hopefully) some concrete benchmarks of its performance? Many thanks, Brian Kingsbury bedk@icsib20.Berkeley.EDU