Scott-
I think what you're pointing out, quite correctly, is that there can be
passive circuit elements which occur on the die but before the
transition to the high impedance level typically used for most of the
receiver. These circuit elements can include series elements as well as
shunt elements. If these series circuit elements are left out of the
network analysis, then the accuracy of the result will necessarily be
compromised.
!->This statement applies regardless whether the analysis is
performed in the time or frequency domain. If you leave circuit
elements out, the accuracy of your analysis will be degraded.<-!
To date, IP vendors have not been inclined to provide equivalent
circuits for their on-die termination networks, with or without passive
series circuit elements; and the most general IBIS models do not
include series circuit elements. If these on-die series circuit
elements are to be included in the network analysis, therefore making
more accurate analyses possible, then one of three things must occur:
1. The IP vendors supply equivalent circuits for their on-die networks.
(Not happening)
2. The IP vendors supply encrypted models (SPICE or otherwise) and we
all use those.
3. IBIS provides a more general mechanism for describing the circuit
behavior of these networks while hiding the internal details of the
network.
In the past, we've proposed on-die S parameters as a relatively simple
and effective solution to this problem.
Mike S.
On 05/28/2010 08:25 AM, Scott McMorrow wrote:
Mike
The interesting thing about the compensation networks being used in
transmitters and receivers today, is that the actual driver or receiver
sits on
a high impedance node formed by mutual inductance cancellation of the
two 50 ohm line and termination ports. A transient pulse response of
the actual circuit
will give the right answer. But if, as I suspect many do,
frequency domain techniques are used to derive the impulse response,
and the network termination and
receiver/driver are considered lumped at the same point, the resulting
impulse
response will be off by 10% and include additional pulse distortion,
which is significant for a die. The
conclusion I made for all impulse response generation
methods is that the receiver/driver needs to be modeled at a minimum as
a
3-port network per SE port, using a high impedance receiver/driver port
whenever necessary, and the impedance deviation from 50 ohms needs to
part of the modeling. Otherwise, the resulting impulse response, or
derived waveforms passed to the AMI DLL will be incorrect. There are
also some numerical issues that need to be addressed for s-parameters
with mixed 50 ohm/10K ohm ports.
regards,
Scott
On 5/28/2010 4:52 AM, Mike Steinberger wrote:
Scott-
Excellent answer. Thanks for the information.
Mike S.
On 05/27/2010 06:25 PM, Scott McMorrow wrote:
I
also
wonder
what
you mean by "large errors". What is your numerical
estimate of the magnitude of such errors? How did you perform that
estimate? What data did you base your estimate on?
Mike S.
Numerical error estimate - approximately 9 to 10%, depending on
compensation inductance
- Lumping the receiver at the termination introduces about an
additional 9% loss in eye height, along with pulse distortion.
Estimate performed using HSPICE transient simulation of receiver circuit
Published data on device input characteristics correlated to
measurements
--
Scott McMorrow
Teraspeed Consulting Group LLC
121 North River Drive
Narragansett, RI 02882
(401) 284-1827 Business
(401) 284-1840 Fax
http://www.teraspeed.com
Teraspeed® is the registered service mark of
Teraspeed Consulting Group LLC
--
Scott McMorrow
Teraspeed Consulting Group LLC
121 North River Drive
Narragansett, RI 02882
(401) 284-1827 Business
(401) 284-1840 Fax
http://www.teraspeed.com
Teraspeed® is the registered service mark of
Teraspeed Consulting Group LLC
|