# RE: [SI-LIST] : Accuracy of HSPICE W-element transmission lines

From: John Carrillo (John.Carrillo@eng.sun.com)
Date: Wed Nov 29 2000 - 10:56:28 PST

Attached is the latest posting on that thread.

- John Carrillo
NSPG/Network Servers

------------- Begin Forwarded Message -------------

From: "Tsuk, Michael" <Michael.Tsuk@compaq.com>
To: "'Haller, Robert'" <rhaller@cereva.com>, "'Larry Miller'"
<ldmiller@nortelnetworks.com>, David Kaiser <david.kaiser@mcdata.com>, Kai Keskinen
<keskinen@nortelnetworks.com>, si-list@silab.Eng.Sun.COM
Cc: "Enver, Ahsan" <Ahsan.Enver@compaq.com>
Subject: RE: [SI-LIST] : Accuracy of HSPICE W-element transmission lines
Date: Wed, 29 Nov 2000 11:29:08 -0500
MIME-Version: 1.0

Bob Haller wrote:

At higher edge rates (< 100ps) we began to see discrepancies. I think these
were limitations in the
Dielectric losses not modeled over frequency, Michael Tsuk, Could answer
this better.

Since Bob mentioned my concerns here, I thought I'd elaborate a bit. My
problem is that the two dielectric loss parameters in the current version of
the WLINE model are insufficient. The DC loss (Go) is not very helpful in
most cases, since our dielectric materials tend to have very low leakage at
DC. The frequency-dependent loss (Gd), coming as it does from a constant
loss tangent, has problems for time domain simulation.

The problem is that the admittance per unit length of the transmission line
in the frequency domain (G + j \omega C) must be the Fourier transform of an
impulse response that is real, causal, and passive. This puts significant
constraints on the behavior of the admittance in the frequency domain. In
particular, the real part must be an even function of frequency, and the
imaginary part must be odd. So the loss tangent can't be constant, since
that would give G proportional to \omega C. For those of you with a physics
background, this is just a consequence of the Kramers-Kronig relation; for
those with filter design experience, imagine trying to design a filter with
constant phase (not linear phase). It is my belief, although I haven't
proven it, is that in trying to match a physical time domain response to a
non-physical frequency response the WLINE model can sometimes get confused.

In any case, with the new table-based model, these problems are in theory
resolved. But the ball is back in our court as to how to generate the
tables properly. Constant loss tangent is unphysical, and yet it seems to
be a reasonable approximation to the behavior of FR4 over a significant
frequency range. I would appreciate an ideas or references anyone has on
this issue.

--
Michael Tsuk
Compaq AlphaServer Product Development
(508) 467-4621
-----Original Message-----
From: Haller, Robert [mailto:rhaller@cereva.com]
Sent: Tuesday, November 28, 2000 5:12 PM
To: 'Larry Miller'; David Kaiser; Kai Keskinen; si-list@silab.eng.sun.com
Subject: RE: [SI-LIST] : Accuracy of HSPICE W-element transmission lines
Larry,
A while ago I did simulation versus measurement comparisons using AVANT!
W-element
and a TDR with high bandwidth FET probes.
(But I used RLGC matrixes generated by Digital's/Compaq's internal Field
solver - SIMPEST) .
I utilized various edge rates 50->500 ps edge rate, a 12 inch transmission
line, and Hspice 99.x
I also acquired and digitized the source and destination waveforms of the
transmission line,
was careful to model the probes in my simulation and use the actual
digitized waveform (PWL)
with the correct output impedance as the input stimulus to SPICE.
A few Key items was to digitize the input waveform going into the W-line and
to Cross-section the
actual Transmission Line you measured.
I was successful in getting W-line simulations to virtually overlay measured
results (far end
of the Transmission line using the W-LINE).

At higher edge rates (< 100ps) we began to see discrepancies. I think these
were limitations in the
Dielectric losses not modeled over frequency, Michael Tsuk, Could answer
this better.

I also found the the simulations were surprising affected be HSPICE OPTIONS;
METHOD, TIMESTEP
Very Long Lines (> 3 Meters had issues - I think this has been fixed)
T-lines were surprising good if you don't care about high frequency effects
and U-Lines were horrible.

Bob Haller
Cereva Networks
Marlboro Ma.
508-486-9660-3365
-----Original Message-----
From: Larry Miller [mailto:ldmiller@nortelnetworks.com]
Sent: Monday, November 27, 2000 1:53 PM
To: David Kaiser; Kai Keskinen; si-list@silab.eng.sun.com
Subject: RE: [SI-LIST] : Accuracy of HSPICE W-element transmission lines
Yes, but have you compared your results (either) with real hardware?
Inquiring minds want to know....
Thanks,
Larry Miller
-----Original Message-----
From:   David Kaiser [SMTP:david.kaiser@mcdata.com]
Sent:   Monday, November 27, 2000 10:17 AM
To:     Keskinen, Kai [KAN:0G15:EXCH]; si-list@silab.eng.sun.com
Subject:        RE: [SI-LIST] : Accuracy of HSPICE W-element transmission
lines
I have compared the W-element with the RLGC matrix from running
APSIM on 100 Ohm diff'l pairs). With a 30ps pulse and 10ps rise times, the
reflected pulse was very symetrical for the W-element, but showed and RC
roll off with APSIM. The APSIM seemed to shown a more realistic reflection
pulse.

David Kaiser
McDATA Corp.
310 Interlocken Pkwy.
Broomfield, CO 80021
(303) 460-4431
davidk@mcdata.com

-----Original Message-----
From: Kai Keskinen [ mailto:keskinen@nortelnetworks.com
<mailto:keskinen@nortelnetworks.com> ]
Sent: Thursday, November 23, 2000 11:19 AM
To: si-list@silab.eng.sun.com
Subject: [SI-LIST] : Accuracy of HSPICE W-element transmission lines
Hello SI-People:
We have now several times encountered differences in results from
other modelling packages and results from HSPICE w-element models for
applications involving fairly long striplines  > 30cm on FR4 and GETEK with
2.5Gbps signals with ~100ps edge rates. The W-element model appears to give
lower loss than the other techniques. Typical track widths are 6-8 mils with
100Ohm loosely coupled differential pairs.
Avant! claims the W-element method is accurate without providing a
paper showing how it was validated.
How does the rest of the SI community feel about the accuracy of the
W-element model?
Kai Keskinen
Equipment and Network Interconnect
Nortel Subsystems and Performance Networks (NSPaN)
(613)-765-3506 (ESN 395)
keskinen@nortelnetworks.com
------------- End Forwarded Message -------------


RE: [SI-LIST] : Accuracy of HSPICE W-element transmission lines
Bob Haller wrote:

At higher edge rates (< 100ps) we began to see discrepancies. I think these were limitations in the
Dielectric losses not modeled over frequency, Michael Tsuk, Could answer this better.

Since Bob mentioned my concerns here, I thought I'd elaborate a bit.  My problem is that the two dielectric loss parameters in the current version of the WLINE model are insufficient.  The DC loss (Go) is not very helpful in most cases, since our dielectric materials tend to have very low leakage at DC.  The frequency-dependent loss (Gd), coming as it does from a constant loss tangent, has problems for time domain simulation.

The problem is that the admittance per unit length of the transmission line in the frequency domain (G + j \omega C) must be the Fourier transform of an impulse response that is real, causal, and passive.  This puts significant constraints on the behavior of the admittance in the frequency domain.  In particular, the real part must be an even function of frequency, and the imaginary part must be odd.  So the loss tangent can't be constant, since that would give G proportional to \omega C.  For those of you with a physics background, this is just a consequence of the Kramers-Kronig relation; for those with filter design experience, imagine trying to design a filter with constant phase (not linear phase).  It is my belief, although I haven't proven it, is that in trying to match a physical time domain response to a non-physical frequency response the WLINE model can sometimes get confused.

In any case, with the new table-based model, these problems are in theory resolved.  But the ball is back in our court as to how to generate the tables properly.  Constant loss tangent is unphysical, and yet it seems to be a reasonable approximation to the behavior of FR4 over a significant frequency range.  I would appreciate an ideas or references anyone has on this issue.

--
Michael Tsuk
Compaq AlphaServer Product Development
(508) 467-4621

-----Original Message-----
From: Haller, Robert [mailto:rhaller@cereva.com]
Sent: Tuesday, November 28, 2000 5:12 PM
To: 'Larry Miller'; David Kaiser; Kai Keskinen; si-list@silab.eng.sun.com
Subject: RE: [SI-LIST] : Accuracy of HSPICE W-element transmission lines

Larry,
A while ago I did simulation versus measurement comparisons using AVANT! W-element
and a TDR with high bandwidth FET probes.
(But I used RLGC matrixes generated by Digital's/Compaq's internal Field solver - SIMPEST) .
I utilized various edge rates 50->500 ps edge rate, a 12 inch transmission line, and Hspice 99.x
I also acquired and digitized the source and destination waveforms of the transmission line,
was careful to model the probes in my simulation and use the actual digitized waveform (PWL)
with the correct output impedance as the input stimulus to SPICE.
A few Key items was to digitize the input waveform going into the W-line and to Cross-section the
actual Transmission Line you measured.
I was successful in getting W-line simulations to virtually overlay measured results (far end
of the Transmission line using the W-LINE).

At higher edge rates (< 100ps) we began to see discrepancies. I think these were limitations in the
Dielectric losses not modeled over frequency, Michael Tsuk, Could answer this better.

I also found the the simulations were surprising affected be HSPICE OPTIONS; METHOD, TIMESTEP
Very Long Lines (> 3 Meters had issues - I think this has been fixed)
T-lines were surprising good if you don't care about high frequency effects and U-Lines were horrible.

Bob Haller
Cereva Networks
Marlboro Ma.
508-486-9660-3365
-----Original Message-----
From: Larry Miller [mailto:ldmiller@nortelnetworks.com]
Sent: Monday, November 27, 2000 1:53 PM
To: David Kaiser; Kai Keskinen; si-list@silab.eng.sun.com
Subject: RE: [SI-LIST] : Accuracy of HSPICE W-element transmission lines

Yes, but have you compared your results (either) with real hardware?

Inquiring minds want to know....

Thanks,

Larry Miller

-----Original Message-----
From:   David Kaiser [SMTP:david.kaiser@mcdata.com]
Sent:   Monday, November 27, 2000 10:17 AM
To:     Keskinen, Kai [KAN:0G15:EXCH]; si-list@silab.eng.sun.com
Subject:        RE: [SI-LIST] : Accuracy of HSPICE W-element transmission lines

I have compared the W-element with the RLGC matrix from running APSIM on 100 Ohm diff'l pairs). With a 30ps pulse and 10ps rise times, the reflected pulse was very symetrical for the W-element, but showed and RC roll off with APSIM. The APSIM seemed to shown a more realistic reflection pulse.

David Kaiser
McDATA Corp.
310 Interlocken Pkwy.
Broomfield, CO 80021
(303) 460-4431
davidk@mcdata.com

-----Original Message-----
From: Kai Keskinen [mailto:keskinen@nortelnetworks.com]
Sent: Thursday, November 23, 2000 11:19 AM
To: si-list@silab.eng.sun.com
Subject: [SI-LIST] : Accuracy of HSPICE W-element transmission lines

Hello SI-People:

We have now several times encountered differences in results from other modelling packages and results from HSPICE w-element models for applications involving fairly long striplines  > 30cm on FR4 and GETEK with 2.5Gbps signals with ~100ps edge rates. The W-element model appears to give lower loss than the other techniques. Typical track widths are 6-8 mils with 100Ohm loosely coupled differential pairs.

Avant! claims the W-element method is accurate without providing a paper showing how it was validated.

How does the rest of the SI community feel about the accuracy of the W-element model?