From: sweir ([email protected])
Date: Tue Nov 02 1999 - 14:12:24 PST
Paul,
Thanks for the article. I think that the article could be better
written. I believe that Mr. Ritchey's basic premise is that for a
significant class of designs where the signals are completely contained
within the PCB, and do not cross any moats, the sole benefit of
differential signaling is improved ground noise immunity at the
receiver. What he appears to leave out of the article are the
parametric
limits involved. I am fairly unhappy with those omissions and the
writing
style in general, but I am relatively sure that Mr. Ritchey knows what
he
is talking about.
Important items either missing, or missing emphasis in the article.:
1. The weak coupling, 10-15% between the two signals in a differential
pair translates to a fundamental requirement to treat each of the two
signals with the same care and feeding as two independent single-ended
signals of like risetime and amplitude. The differential nature of the
signals within the PCB is not going to significantly help EMI versus the
two signals taken separately. So long as the signals remain confined to
the PCB.
2. That only to the extent that some additional timing uncertainty due
to
variations in and Tr/Tf are acceptable, the two trace lengths may vary
from one another.
3. Than in no case should the trace lengths be allowed to vary by more
than the lesser of the minimum Tr or Tf.
4. That the high sensitivity of the differential receiver comes at a
price. Any differential noise on the pair from an aggressor will shift
the
switching point in time, or may even cause a double-transition in the
receiving output. This is a very good reason to keep the pair together,
and away from aggressors such as TTL/CMOS. This is independent of
whether
lead lengths are matched or not.
5. Since the signals will each independently reflect mostly from the
nearest ground plane, matching vias between the ground plane(s) and each
of
the signals is more significant than matching vias between the two
signals
when switching layers.
I disagree with his characterization of an emitter-coupled differential
amplifier as a "current switch". A diff amp is a diff amp. It
definitely
has a linear region, and the output slew rate definitely depends on the
amount of overdrive at the input. When one of the wavefronts arrive at
the
receiver delayed, the switching point moves up or down referenced to
ground. Assuming perfectly trapezoidal source signals, the crossover
point
in time is delayed:
Tdelay = ( Tslew + Toffset ) / 2
Where Tdelay is the crossover point at the receiver, Toffset is the time
offset in the signal pairs, and Tslew is the slewing time of either
signal,
presumed to be equal for both.
If the signals are perfectly trapezoidal, and the amplifier is perfectly
linear throughout its common mode region, then Tdelay is constant for a
given Toffset. And, Toffset may be varied up to, but no more than
Tslew. ( If the value approximates or exceeds Tslew, timing can become
very erratic. ) However, neither assumption holds absolutely in
real-life. The result is some loss of symmetry, and introduction of
jitter.
How bad can any of this be? Given that Mr. Ritchey advocates limiting
Toffset to 1/3 Tslew, the Tdelay will be no more than Tslew / 6. The
additional jitter and loss of symmetry will be some fraction of this
number
due to non-linearities in the source signals, and the receiving
amplifier
response. The switching point voltage moves by ( Vhi - Vlo ) / 6 or
less. The signal crossover point now lies between 33% and 67% as
compared
to an ideal of 50%. It would take a good number of very careful
measurements to determine what the delay sensitivity to the common mode
switching point is. Perhaps someone would like to do a paper on this.
The total distortion is likely to come-out quite small. Whether or not
it
matters depends on the application circuit. For instance, a Channel
Link
type device at the upper clock rates has timing margins in the 10's of
pS. In such a case, the recommended technique could be a real problem
if
taken to the limits recommended.
As I have said before, I disagree with his categoric statement that the
total impedance is just the single-ended impedance divided by two. That
statement ignores the mutual coupling of 10-15%, so I expect a 5-8%
calculation error.
In the end, I believe that the article could have been written better.
For
a number of cases, I believe the article's intent is valid within the
confines of a PCB. I do not agree at all with this method for any
signals
which leave a box, as the offset dramatically increases common noise
which
is likely to become an emissions nightmare.
Regards,
Steve.
At 08:04 AM 10/29/99 -0400, you wrote:
> It is also stated in this article that critically matched lengths is
>unnecessary, however you MUSt take into account your noise margins(parallel
>aggressors). It states that differential transmissions only need about 15mV
>difference to switch.
> Assuming no noise/ crosstalk and a 300pSec rise time in FR-4(180
>pSec/In), the pairs length could be mismatched by as much as 1.6
>inches(Tr/Td) and still function. So a 500 mil difference in line length
>would be acceptable. What he is really stating is that if you need to route
>one of the traces around a through hole, just do it.
>
> It only shows an example for ECL and LVDS Logic in the acrticle.
>
>Regards,
>
>Paul Denomme
>
>
> > -----Original Message-----
> > From: S. Weir [SMTP:[email protected]]
> > Sent: Thursday, October 28, 1999 6:04 PM
> > To: [email protected]
> > Subject: RE: [SI-LIST] : Comments from your SI seminar (SendII)`
> >
> > Paul,
> >
> > Are you assuming:
> >
> > 1. The traces have been specified to remain at a constant separation?
> > 2. The trace pair will have a minimum separation from parallel aggressors?
> > 3. The traces will have matched lengths?
> >
> > I am not sure how such things can be assumed and make it to the physical
> > design. If the trace lengths don't match, the signal will have lots of
> > common mode for fast enough edges. If there is not enough separation from
> >
> > parallel aggressors, then aggressors can inject differential mode
> > noise. If the traces do not maintain a constant separation, the impedance
> >
> > will vary by about half the variation in the coupling coefficient.
> >
> > Regards,
> >
> >
> > Steve.
> >
> > At 03:30 PM 10/28/1999 -0400, you wrote:
> > > I have read an article recently that states that the use of
> > >specifying the differential impedance of two traces on a circuit board is
> > >unnecessary. The only thing you need to worry about is the individual
> > trace
> > >impedance. If you need a differential impedance for two lines to be 100
> > >ohms, just use two 50 ohm lines rather than using two signals whose
> > >differential impedance is 100 ohms. Also when connecting a 110 ohm
> > twisted
> > >pair to PCB you should just connect it to two 55 ohm traces to achieve
> > the
> > >110 ohm differential impedance. I have done enough research to draw my
> > own
> > >conclusions, but I would like to get the reaction from people in this
> > forum
> > >regarding this issue.
> > >
> > >Thank you,
> > >
> > >Paul Denomme
> > >Viasystems Inc.
> > >
> > >
> > >
**** To unsubscribe from si-list: send e-mail to [email protected]. In the BODY of message put: UNSUBSCRIBE si-list, for more help, put HELP. si-list archives are accessible at http://www.qsl.net/wb6tpu/si-list ****
This archive was generated by hypermail 2b29 : Tue Feb 29 2000 - 11:39:32 PST