Re: Frequency Dependent and Lossy

From: Ted Creedon <ted_creedon@mentorg.com>
Date: Wed Nov 24 1999 - 09:06:09 PST

You are right on.

Just a 400 mhz connector can cause problems due to inaccurate modeling therefore
inaccurate specification by the customer therefore lack of process control by the
vendor. And this is in addition to the Z0 change at the connector.

Two things will happen 1. the designers will get smarter and stop using high speed
connectors thus avoiding the problem and 2. Ibis will eventually figure out how to
model frequency dependent devices.

Sets of S parameters for N ports are the only comprehensive solution now because the
only accurate information available is from instruments or 3D field solvers. However
the condensation of the information into something useable (the S parameter sets can
be viewed as a (N-1)*ntables dimensioned surface where N=# of ports, ntables=no of
frequency points for each Nport set of S parameters). We're running into this
already for via modeling.

The problem of characterizing e' + je'' which vary with frequency is a area of
active interest. Particularly since many designers choose to create this problem by
using FR4 rather than a good dielectric like Duroid or Kapton. Which is worse
because for FR4 boards e' and je'' are anisotropic - it varies with the fiberglass
mat orientation in XY and and may also vary in the Z direction. Don't forget you're
trying to control the electrical length to less than 100ps and keep the rise time
out of the quadratic range (RC line rather than LC).

As an example a large digital system could be flip chipped on a MCM - no PCI busses,
AGP busses, ISA busses or memory connectors with a microwave launch to the
keyboard/mouse and display. And satellite or fiber.
If one stops and thinks of the advantages in miniaturizing, lowering manufacturing
cost and simplifying simulation we might design ourselves out of a job.

ted

apanella@molex.com wrote:

> First,
>
> To the entire group... I am sorry that I missed the last meeting. I was fully
> planning on being there... had an unfortunate incident that came up that need
> my immediate attention...
>
> I will plan on being at the next meeting.
>
> However, in the spirit of helping... I would like to address a few of the
> comments from the meeting notes:
>
> This email topic:
>
> "Stephen Peters and Arpad Muranyi asked several clarifying questions on
> whether the (frequency dependent) conductance matrix was included, and Kellee
> responded that it was not. Furthermore he stated that the decision to
> postpone losses was made after consulting with a number of internal technical
> experts."
>
> In my mind there are two issues built into this statement:
>
> 1. Frequency dependence: At this time a "FREQUENCY DEPENDENT COMPONENT" is not
> built into the specification... However... CASCADED MATRICES are built in to
> the specification. The cascaded matrix allows the model to represent
> asymetrical or symetrical filtering effects that are typically found in
> interconnect structures with multiple discontinuities (i.e. allows you to
> represent a pole/zero as they actually occur as the energy passes through the
> length of the interconnect). Thus representing each discontinuity on it's own..
> in a series and/or stub configurations. As such, it is quite reasonable to
> expect that by using a cascaded model, it is possible to extend the operation
> point of a model to the limits of the performance limits of the interconnect
> being modeled.
>
> With BERKELEY SPICE models (no vendor specifi SPICE extensions required) , using
> a cascaded modeling approach... today, a 6GHz 3dB bandwidth is a typical
> modeling requirement (9GHz. is also possible). Yes this can be confirmed to
> both frequency domain and time domain lab measurements.
>
> Further, I would also suggest that a frequency dependent component is really
> something more than just a swept series resistance that varies as a function of
> frequncy in a single lumped model. As such, to do a true frequency dependent
> model, something more like a behavioral model (s-parameters anyone?) would be
> the most appropriate method of representation.
>
> 2. Conductance matrix: At this time a conductance matrix as such is not built
> in to the specification. I think that when this effect comes into play (I do
> not have supporting data, but my guess that conductance becomes significant at
> frequencies greater 8 GHz. ....at least in devices as short as connectors) that
> the "s-parameter" solution would be the best solution to correctly implement
> dielectric loss effects.
>
> So why noy implement s- parameters now
> * Model complexity.... the connector matrix models that are setup in the
> proposed connector specification can be in the 100's to 1000's of lines. I've
> seen (and created) mulitport s-parameter models that are 100x the size. Of
> course, this depends on the frequncy step size and frequncy range of the model.
> * We really wanted to incorporate an unlimited pin size connector using a
> "swath" as proposed in the specification. As the s-parameter matrix is not a
> "single value per coupling / self node" (like , for instance an inductance
> matrix would be)... I had no idea on how to handle the complexity of a swathed
> s-parameter.
> * Many simulators that use IBIS models are time domain based... the complexity
> of adding s-parameters may be too difficult.
> * The proverbial , We approached the proposed specification with a "Walk before
> you run philosophy".
> * We also wanted to focus on the connector models that are primarily available
> today... and be able to implement those as soon as possible into a new format.
>
> _gus: 630-969-4617
> apanella@molex.com
Received on Wed Nov 24 09:07:11 1999

This archive was generated by hypermail 2.1.8 : Fri Jun 03 2011 - 09:52:30 PDT