RE: about ibs model's ac analysis

Chris,
 
Answers to your questions:
 
1) Yes, the pole/zero or the transfer function does change significantly
due to the
DC bias point. One of the most obvious change is that the term which
represents
the resistance corresponds to the slope of the IV curve. Near the origin I
can get
a few Ohms, in the saturation region I can get hundreds of Ohms. Reading
Fred's
response (and the manual) I agree that this is due to the difference between
small
signal vs. large signal model of the same MOSFET in (H)SPICE.
 
2) Regarding the equivalent circuit, there are multiple ways to make a
voltage
dependent model. This is how I did it: First I generated the Z11 data for
a MOSFET.
at various operating points. Then I ran an optimization routine to fit
either R and C
values or coefficients for a transfer function (using the Laplace element)
to get the
same Z11 response. Then I curve fitted these parameters and built an RC
circuit
and a Laplace source using the tansfer function with these voltage dependent
parameters.
 (Yes the Laplace source works this way if you don't have more than three
terms in it).
Then I overlayed the original MOSFET with the RC circuit and the Laplace
source's AC
response to make sure I get the same Z11 response with respect to DC bias
changes.
So far so good. Then I ran the same three circuits in .TRAN mode to see if
the
natural responses will be the same. The RC and Laplace matches with each
other, but
they don't match the MOSFET. I kind of knew that the reason is as explained
above,
I just wanted to get other's views also. (By the way, this discrepancy onyl
exists when
the original device is a MOSFET. If I used an RC circuit as the original
element, the
process that I just described ends up giving the same results after all this
rigamour).
 
Now, I understand that the small and large signal models of transistors are
different.
But my concern is that those who worship SPICE models may be in for a big
surprise
if they run all kinds of frequency domain simulations (using .AC) thinking
that they are
solving resonance problems and all of the sudden their fixed circuit will
not work in the
time domain!
 
Arpad
============================================================
 -----Original Message-----
From: Chris Rokusek [mailto:crokusek@innoveda.com]
Sent: Tuesday, March 27, 2001 10:34 AM
To: ibis-users@eda.org
Subject: RE: about ibs model's ac analysis

Arpad,
 
I am interested in your frequency domain work here...
 
Do the Laplace pole/zero locations vary significantly based on the DC bias
point? If yes, then it seems you are attempting to compare a "small-signal"
equivalent with a "large-signal".
 
I am reading that you created an equivalent circuit that attempts takes into
account these different bias's, however I didn't think that was possible in
general. Is there a standard procedure to synthesize a circuit given
multiple small signal equivalents? The reason this seems difficult is
because in the time domain, the "small-singnal" model would need to be
changing between the biases in a (perhaps complex) manner.
 
Basically I am asking how you "morphed" the small-signal equivalent over
time to account for biases.
 
Chris
 

-----Original Message-----
From: Muranyi, Arpad [mailto:arpad.muranyi@intel.com]
Sent: Tuesday, March 27, 2001 8:55 AM
To: 'wqzhang@avanticorp.com'; ibis-users@eda.org
Subject: RE: about ibs model's ac analysis

Wenqing,
 
Your observation is true, but not totally. It is true that the
IV curves of IBIS models are made with .DC or .TRAN mode. This
can only provide a frequency independent impedance (or R) for a
given operating point (voltage). However, IBIS models also have
a parameter called C_comp, which is a primitive way of introducing
some AC characteristics. This C_comp is typically in parallel
with the IV curve's non linear impedance (R), so the overall effect
is basically a parallel RC circuit. With the new feature of C_comp
being split into four parts the situation can get a little more
complicated, but in an AC analysis sense this is still pretty simple.
 
Now, I agree that this simple RC circuit is not very sophisticated
for a frequency domain analysis. However, in this initial
implementation the emphasis should not be on the buffer's AC accuracy.
The IBIS open forum is already working on improvements to the
specification which will enable us to make more accurate AC models
also (IBIS-X) in the future.
 
Currently, frequency domain analysis in Signal Integrity Engineering
of digital products (such as computer boards) is used primarily to
detect problems with the characteristics of the interconnects. In
some cases it is desirable to include the driving or terminating
impedances of the buffers and receivers in the AC analysis. When
we have nothing but IBIS models available, it would be useful to be
able to this with the B-elements. However crude it may be, it is
still better than nothing.
 
On the other hand, since I am working on this issue now, I have discovered
that a MOSFET transistor is actually not all that complicated in terms
of frequency domain analysis. I can replace the transistor with a
one-port equivalent (as if I was looking into it from the die pad)
using a transfer function with two or three terms in the (numerator
and) denominator or with a few poles. (I tried this with the pole/zero,
and the Laplace elements in HSPICE).
 
I found a problem, though. I ran a pole/zero analysis to find the
poles (and zeroes) of such a one port of a MOSFET. (It didn't have
zeroes). I ran this at multiple DC bias points to be able to include
voltage dependencies. Then I made an equivalent circuit using
voltage dependent resistors and capacitors, and ran a time domain
simulation to see whether the step or natural response of this
equivalent circuit will give me the same result as the original
MOSFET. Well, it DOESN'T! I tried to look for mistakes etc...,
but the only conclusion I get is that the .AC and the .TRAN models
of the MOSFET are not the same in HSPICE, and that is where the
difference comes from. I would like someone interested in this
to verify it for me.
 
Now, the philosophical question is this: The above story tells us
that the transfer function of a MOSFET is different for the AC and
TRAN modes in HSPCIE. Should they not be the same? If they are
different, my AC analysis will not give me the same picture of what
my system does compared with a TRAN analysis. Is that right?
 
Arpad Muranyi
Intel Corporation
=================================================================
 

 
Received on Tue Mar 27 17:13:36 2001