Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!tut.cis.ohio-state.edu!rutgers!mcnc!thorin!unc!jp
From: jp@unc.cs.unc.edu (John Poulton)
Newsgroups: comp.lsi
Subject: CAzM circuit simulator
Message-ID: <7737@thorin.cs.unc.edu>
Date: 13 Apr 89 17:43:59 GMT
Sender: news@thorin.cs.unc.edu
Distribution: usa
Lines: 43


There has been some discussion lately in this newsgroup about circuit
simulators that might be "better" than SPICE by some measure.  I'd
like to contribute our recent experience with a simulator called CAzM,
developed by a collaboration between Duke University CS (Don Erdman,
Don Rose) and the Microelectronics Center of North Carolina (Gary
Nifong, Ravi Subrahmanyan).  We've been using this simulator to develop
a pair of CMOS IC's with fairly aggressive speed and density requirements
and have found CAzM to be extremely useful and quite robust.  Essentially
every one of the hundred or so cells in these designs has been 
thoroughly simulated, many across a variety of process, voltage, and 
temperature conditions.  If we had used SPICE, I doubt if we could have 
found either the patience or the machine cycles to finish this task.

CAzM takes as input a standard SPICE-like circuit file (with alpha-
numeric node names, of course) and its own simple simulation control
commands, similar enough to SPICE that translation is easy.  It models 
all the usual linear circuit elements and has employed the mos2 model 
for some time.  Recently we have added and tested the mos3 model; MCNC 
will shortly release a version with bsim models and models for bipolar 
transistors.  As we found by experience, the program is well constructed, 
and it's possible to add your own functions with virtually no risk of 
breaking already-existing ones.  Except for certain numerical routines, 
CAzM is written entirely in C; it'll run on Sun and VAX UNIX platforms.

CAzM is MUCH faster than SPICE for mosfet circuits.  It uses the mos2/3/4
models to build tables of drain currents for each device size encountered,
then, during transient analyses, interpolates currents from these tables,
rather than computing them from the model for each time step.  The result
is that CAzM is about 10-60x faster than SPICE (by our comparisons), while
producing essentially identical simulation results.  It uses some advanced 
numerical techniques that make it very robust against convergence problems.  
We have only once or twice encountered the famous "time-step too small" 
problem with this tool, and I believe the MCNC/Duke team is currently 
ironing out the few remaining bugs that occasionally cause these problems.

We'd be very interested in seeing this tool more widely distributed; more 
designer feedback usually produces better design tools.  CAzM is available
to universities at some nominal charge and is distributed by MCNC.  If 
you're interested, get in touch with Steve Kenkel (kenkel@mcnc.org).   

John Poulton
jp@cs.unc.edu