[SI-LIST] Re: package SSN model accuracy requirements, now Behavioral Modeling

  • From: steve weir <weirsi@xxxxxxxxxx>
  • To: "lgreen" <lgreen22@xxxxxxxxxxxxxx>, <kumarchi@xxxxxxxxx>, <si-list@xxxxxxxxxxxxx>
  • Date: Sat, 19 Mar 2005 20:15:11 -0800

Dr. Green, I think that the irony is that the problem with macromodels 
found in Dr. Lauritzen's work is that the available library lacked behavior 
sufficiently close to reality for the circuit structures he wished to 
model.  That is not really any different than Chris' criticism of IBIS.

Regards,


Steve.
At 07:58 PM 3/19/2005 -0800, lgreen wrote:
>Hi, Steve,
>
>The real sticking point is that IBIS models is what I/O designers provide,
>and IBIS does not model tightly coupled diff pairs.  If you get a SPICE
>netlist, you are better off, of course.  But even SPICE models are not
>foolproof - back when I used to design I/O buffers, I remember seeing some
>flaky BSIM models from some of the foundries we dealt with.  ("Little"
>things like discontinuities in I vs Vds.)
>
>The links under Peter Lauritzen's page show a number of devices and circuits
>where no macromodels could match the transient circuit characteristics,
>while a well-built behavioral model provided a better match to data (i.e.
>accuracy) and ran faster.
>
>I guess the bottom line is to do whatever modeling it takes to get the
>design out the door, while also verifying models whenever possible.
>
>- Lynne
>
>-----Original Message-----
>From: si-list-bounce@xxxxxxxxxxxxx [mailto:si-list-bounce@xxxxxxxxxxxxx] On
>Behalf Of steve weir
>Sent: Saturday, March 19, 2005 7:29 PM
>To: lgreen22@xxxxxxxxxxxxxx; kumarchi@xxxxxxxxx; si-list@xxxxxxxxxxxxx
>Subject: [SI-LIST] Re: package SSN model accuracy requirements, now
>Behavioral Modeling
>
>Dr. Green, the trick of course is getting sufficient accuracy into the
>model, no matter the form.  I think that particular sticking point is what
>brings us back to Chris' position that such models are presently difficult
>to obtain in a form other than a SPICE netlist.
>
>Regards,
>
>
>Steve.
>
>
>At 07:17 PM 3/19/2005 -0800, lgreen wrote:
> >You have hit the nail on the head!
> >
> >Macromodels are not required to be implemented as current sources (one
> >could, for example, use controlled V sources).  The simulator must still
> >apply KCL and KVL to set up and solve the circuit matrix.
> >
> >The "philosophy" question is when to write behavioral models (equations)
>and
> >when to write macromodels (adding components).  The classic SPICE 2g6
>matrix
> >solution time increased as N_cubed, so if a macromodel added nodes, it
>could
> >greatly increase run time.
> >
> >I have seen comparisons for power devices, with an order of magnitude
>faster
> >run times for behavioral models in simple circuits with better accuracy
>than
> >the macromodel.
> >http://www.ee.washington.edu/people/faculty/lauritzen/
> >  I personally would love to see similar comparisons between the two
> >approaches for circuits containing several I/O buffers.  Maybe someone will
> >present these at future IBIS summits.
> >
> >- Lynne
> >
> >
> >-----Original Message-----
> >From: si-list-bounce@xxxxxxxxxxxxx [mailto:si-list-bounce@xxxxxxxxxxxxx] On
> >Behalf Of C. Kumar
> >Sent: Saturday, March 19, 2005 3:50 AM
> >To: lgreen22@xxxxxxxxxxxxxx; donaldt@xxxxxxxxxxx; 'Syed Huq'
> >Cc: si-list@xxxxxxxxxxxxx
> >Subject: [SI-LIST] Re: package SSN model accuracy requirements, now
> >Behavioral Modeling
> >
> >lynn:
> >
> >Your definition makes sense. If we go by that we can
> >the say the following about "macromodels" written in
> >spice
> >
> >1. all spice elements are behavioral elements. They
> >fundamentally model the current flow; all spice
> >elements are some form of current sources. As you
> >point out spice does not have "pure" equations.
> >Everything has to be behavioral components connected
> >in a netowork and spice imposes conservation at the
> >nodes.
> >
> >2. For our purpose we can say the purpose of
> >macromodels
> >is
> >  a. compactly reproduce the behavior of a complex
> >device  using a collection of current sources (our
> >behavioral primitives i.e all the 'elements' available
> >in a given spice simulator).
> >
> >  b. like in your example op amp, include "non-ideal"
> >effects like that may be contained in a post layout
> >model.
> >
> >--- lgreen <lgreen22@xxxxxxxxxxxxxx> wrote:
> >
> > > Hi, Donald and Syed,
> > >
> > > In academia, a distinction is made between the two
> > > modeling approaches,
> > > although you are correct that the distinction is
> > > often fuzzy in industry.
> > >
> > > "Macromodeling" involves adding components to a
> > > model.  A textbook example
> > > is adding components to an ideal op amp to add
> > > effects like input offset
> > > voltage, input bias current, output impedance, poles
> > > and zeroes, etc.  A
> > > macromodel's added components could be passive or
> > > active, linear or
> > > non-linear.  In this sense, all netlist-based tools
> > > (such as SPICE) can do
> > > macromodeling.
> > >
> > > "Behavioral modeling" involves creating a model from
> > > performance
> > > characteristics rather than topology, usually using
> > > equations or data
> > > tables.   IBIS 3.2 is the classic example, but there
> > > are hundreds of
> > > AMS-style models that have been created over the
> > > past twenty years.
> > >
> > > There is also a "meet in the middle" approach, where
> > > equation-based V and I
> > > sources are added to a model.  The use of a V or I
> > > component makes it
> > > macromodeling (in the academic sense of the word).
> > > SPICE 3f5 seems to be a
> > > common subset of programmable sources for general
> > > purpose analog simulators
> > > (which means it is the most limited).
> > >
> > > There are advantages to both approaches.
> > > Macromodeling is easier to do if
> > > there is an existing model (such as an ideal op amp
> > > or IBIS buffer) that is
> > > almost what you need.  Behavioral models can
> > > simulate faster (factors of 10
> > > to 200, depending on how the model is coded), since
> > > they do not add
> > > extraneous components or nodes to the simulation
> > > netlist.
> > >
> > > Best regards,
> > > Lynne
> > >
> > >
> > > "IBIS training when you need it, where you need it."
> > >
> > > Dr. Lynne Green
> > > Green Streak Programs
> > > http://www.greenstreakprograms.com
> > > 425-788-0412
> > > lgreen22@xxxxxxxxxxxxxx
> > >
> > >
> > > -----Original Message-----
> > > From: si-list-bounce@xxxxxxxxxxxxx
> > > [mailto:si-list-bounce@xxxxxxxxxxxxx] On
> > > Behalf Of Donald Telian
> > > Sent: Friday, March 18, 2005 9:49 AM
> > > To: Syed Huq
> > > Cc: si-list@xxxxxxxxxxxxx
> > > Subject: [SI-LIST] Re: package SSN model accuracy
> > > requirements, now
> > > Behavioral Modeling
> > >
> > > Syed,
> > >
> > > Thanks for helping us refine terminology.
> > >
> > > Let's be careful not to mix up concepts with tools
> > > and languages.  While
> > > it seems you want to confine "macromodeling" to
> > > something only Cadence
> > > tools do, I believe that if you look around you will
> > > find that
> > >
> > > macromodeling =3D=3D behavioral modeling
> > >
> > > ...and the two are used interchangeably within
> > > industry and academia.
> > > If you reference every one of the research papers I
> > > give links to on
> > > slide 22 - as well as the quote from Franzon - you
> > > will note that they
> > > all use "macromodeling" in both the titles and paper
> > > texts to describe
> > > their work in "behavioral modeling".  They use the
> > > term "macromodeling"
> > > to refer to various implementations spanning
> > > equation-based Hspice
> > > models, VHDL-AMS models in ICX, SPICE models in IBM
> > > PowerSPICE, etc.
> > > And yes, "macromodeling" can also be done in
> > > DML/Espice, Pspice,
> > > Berkeley SPICE, Verilog-AMS, and a variety of other
> > > tools/languages.
> > >
> > > Regarding your comments about Cadence tools, I'll
> > > acknowledge that it
> > > takes a fair amount of skill and craftsmanship to
> > > construct effective
> > > macromodels.  However, I've yet to find a device
> > > that can not be
> > > behaviorally modeled in the environment.  Kumar and
> > > others have shown
> > > considerable skill in adding things like
> > > time-controlled sources, nth
> > > derivative processing, multi-dimensional table-based
> > > controlled sources,
> > > and such into our tools to make this possible.
> > > Thanks to their
> > > contributions, users have had a number of published
> > > successes (slide 21)
> > > and implementing your BIRD95 ideas would be a snap.
> > > If your team is
> > > struggling with other particular implementations,
> > > Cadence can provide
> > > you assistance in a variety of ways and would be
> > > happy to do so.
> > >
> > > During the years while it's simpler for IC companies
> > > to encrypt and ship
> > > their transistor layout-derived model, and model
> > > users continue to ask
> > > for faster behavioral solutions, I think it's
> > > important that we keep an
> > > open mind regarding the options before us.  Working
> > > together, let's
> > > encourage those working on the various tools and
> > > languages as new
> > > solutions emerge and keep the comments on the
> > > positive side.  It's a
> > > challenging problem to solve.
> > >
> > > Donald
> > >
> > >
> > > >-----Original Message-----
> > > >From: Syed Huq [mailto:shuq@xxxxxxxxx]=20
> > > >Sent: Wednesday, March 16, 2005 2:44 PM
> > > >To: Donald Telian
> > > >Cc: si-list@xxxxxxxxxxxxx
> > > >Subject: Re: [SI-LIST] Re: package SSN model
> > > accuracy=20
> > > >requirements, nowBehavioral Modeling
> > > >
> > > >Donald,
> > > >
> > > >Certainly our definitions are different. What you
> > > described is=20
> > > >what I would call behavioral modeling. There is a
> > > difference=20
> > > >and let me explain with an example:
> > > >
> > > >Let's pick your tool. SpecctraQuest uses this DML
> > > scheme(which=20
> > > >I call macromodeling), in practical applications
> > > such as ours,=20
> > > >we see the tool choking in many aspects in it's use
> > > of macro=20
> > > >modeling. Hence I would like to stay far away from
> > > such an=20
> > > >approach(macromodeling I mean). Maybe version 20.x
> > > would have=20
> > > >it all sovled. I am sure there are 10s of other
> > > application=20
> > > >where is excels quite well too.
> > > >
> > > >On the other side, behavioral modeling (atleast
> > > again in our
> > > >applications) has already proven it's capabilities
> > > many times=20
> > > >over and I am more than open to explore further
> > > with tools to=20
> > > >see how BIRD95 could be implemented thru
> > > "behavioral modeling".
> > > >
> > > >Syed

The weirsp@xxxxxxxxxx e-mail address will terminate March 31, 2005.
Please update your address book with weirsi@xxxxxxxxxx


------------------------------------------------------------------
To unsubscribe from si-list:
si-list-request@xxxxxxxxxxxxx with 'unsubscribe' in the Subject field

or to administer your membership from a web page, go to:
//www.freelists.org/webpage/si-list

For help:
si-list-request@xxxxxxxxxxxxx with 'help' in the Subject field

List FAQ wiki page is located at:
                http://si-list.org/wiki/wiki.pl?Si-List_FAQ

List technical documents are available at:
                http://www.si-list.org

List archives are viewable at:     
                //www.freelists.org/archives/si-list
or at our remote archives:
                http://groups.yahoo.com/group/si-list/messages
Old (prior to June 6, 2001) list archives are viewable at:
                http://www.qsl.net/wb6tpu

Other related posts:

  1. Home
  2. si-list
  3. Archive
  4. 03-2005