[SI-LIST] : Re: Is feedback really EVIL?

D. C. Sessions (dc.sessions@vlsi.com)
Thu, 14 May 1998 08:29:37 -0700

Charles Hill wrote:
> DC Sessions wrote:
> >Eeeeyeew! Feedback! Feedback is evil. Bad driver, BAD.
> >Go to bed without stability. Naughty, naughty, naughty.
> >Well, OK, that's overstating it. Sometimes feedback is
> >not only useful but necessary; it still makes behavioral
> >modelling (a la IBIS) pretty well impossible. IMNSHO, this
> >is more annoying than crippling since feedback has other
> >properties that limit its usefulness in applications where
> >signal integrity is a major concern, so giving up behavioral
> >modelling of feedback-controlled drivers is (again, IMO) no
> >great loss.
> I guess one man's meat is another man's poison.
> Usually I'm more interested in solving an SI problem, than ensuring that it
> is easy to model. So we agree there.
> So I have several questions:
> 1. Feedback is a very useful technique in many areas of EE. How widely is
> it used in SI? What are the problems with using feedback?

Can't speak for everyone, but in my own experience: not much.
The main exception seems to be active clamping, and even that
is somewhat controversial. All of the drivers I've seen which
try to drive observe the load and adjust accordingly have been
less than brilliant successes.

> 2. Why is a feedback system difficult (pretty well impossible) to model
> with IBIS? I know we had lots of problems achieving accurate modeling of a
> feedback driver in SPICE.

The whole point of IBIS is that it's a simple behavioral model of
a component. The output is a simple V/I function, stuff like that.
The original IBIS objectives even included that it only contain
characteristics that were observable externally (which is why the
v/t curves are the way they are rather than some less ambiguous
internal function).

Feedback means that the model *doesn't* run open-loop; its behavior
both includes modelling internal operations but also depends on the
state of the rest of the system. This tremendously complicates the
modelling language, as well as slowing down the simulators by loads.

Caveat: some of this is second-hand; I don't write SI software and
only pass on what I've learned from the pros who do. If anyone
REALLY wants to know how ugly feedback makes the modelling problem
they can take it up on the ibis@eia.org mailing list.

> 3. What are the properties of feedback that limit its usefulness in
> applications where signal integrity is a major concern?

Speed and stability. Speed because many SI events occur faster than
a feedback path can react, so the circuit is effectively running
open-loop anyway. Stability because SI events occur in a reactive-load
environment with an infinite number of poles, none of them knowable
in advance. That makes stabilizing the feedback system a bear, and
the most straightforward solution is to slow it 'way down (see first

> DC Sessions also wrote:
> >IBIS can't really deal
> >with JEDEC flexible-impedance drivers, and I'm writing THOSE into
> >the IEEE 1394b spec. Great fun. Still, we need to keep in mind that
> >the more exotic variants that you and I deal with are of very narrow
> >application.
> 4. What is a flexible impedance driver, and how does it work?

A flexible-impedance driver is one which has a drive impedance
which can be adjusted to match its load. JEDEC has a standard
in the works which addresses some of the issues involved. The
obvious method is to have a reference pin on the IC which is
tied to ground via a resistor with five times the desired
impedance. This allows well-controlled reflected-wave switching
without the component cost, board crowding, and SI compromises
of external resistors.

D. C. Sessions