[ibis-macro] Re: CTLE and Wave Shape filtering?
- From: "Dmitriev-Zdorov, Vladimir" <vladimir_dmitriev-zdorov@xxxxxxxxxx>
- To: <msteinb@xxxxxxxxxx>, "Scott McMorrow" <scott@xxxxxxxxxxxxx>
- Date: Fri, 15 Oct 2010 11:51:30 -0600
Mike,
>This brings up another useful general principle: Nonlinear effects are
more important for low loss channels (e.g., < 5 dB at r/2) than they are
for higher loss channels (e.g., > 5 dB at r/2). Consider that the first
order effect of a nonlinearity is to generate harmonics of the input
signal, thus increasing the spectral density at higher frequencies. For
a low loss channel, these higher frequency spectral components will be
evident at the receiver input and will therefore affect the receiver's
behavior. This is usually the case for a parallel interface (e.g.,
DDR2), for example. For a lossy channel, however, the loss increases
with frequency and so the channel attenuates the higher frequency
spectral components (including those generated by any nonlinearities)
more than it does the lower frequency spectral components (especially
those generated by the linear response).
This type of reasoning works for periodic signals where spectrum is a
set of harmonic amplitudes and the DC component of the signal is
constant. We indeed can expect considerable attenuation of high
frequency harmonics be the input strictly periodic. Unfortunately, this
is essentially not the case, primarily because periodic signals cannot
transmit useful information. With arbitrary a-periodic input the
spectrum becomes continuous; instead of a constant DC component we
observe slowly changing operation point (with frequency well below the
base frequency) that follows the floating imbalance of the signal (an
average number of zeros and ones within a certain time window as
simulation progresses). The effect is not about 'higher' harmonics, but
mostly about 'modulation' of magnitudes near the base frequency while
the operation point wanders along IV curves. Of course, common mode
effects would add to this a lot.
Vladimir
-----Original Message-----
From: ibis-macro-bounce@xxxxxxxxxxxxx
[mailto:ibis-macro-bounce@xxxxxxxxxxxxx] On Behalf Of Mike Steinberger
Sent: Friday, October 15, 2010 10:39 AM
To: Scott McMorrow
Cc: Scott McMorrow for Jim Bell; ibis-macro@xxxxxxxxxxxxx
Subject: [ibis-macro] Re: CTLE and Wave Shape filtering?
Scott-
This conversation is useful in that it addresses the assumptions that
have been the basis for algorithmic modeling from the very beginning. I
think it's useful to state those assumptions clearly and examine the
conditions under which they're valid.
Your statement
Reflection behavior for the Tx silicon will be incorrect, since
transmitter complex impedance is dependent upon not only the termination
resistance, but any pulse shaping that is performed in the termination
network.
only makes sense if you assume that the output impedance is a function
of the output voltage. While this assumption is certainly true for a
single ended driver, it turns out to be a second order effect for a
balanced differential driver. We have demonstrated this with SPICE
models for a number of different differential drivers from a number of
different IP suppliers. The nonlinear effects on reflection coefficient
are there, but they're so small that you really have to look for them to
find them. In the case we examined in greatest detail, the effect at the
driver output was less than 1% of the driver amplitude.
This brings up another useful general principle: Nonlinear effects are
more important for low loss channels (e.g., < 5 dB at r/2) than they are
for higher loss channels (e.g., > 5 dB at r/2). Consider that the first
order effect of a nonlinearity is to generate harmonics of the input
signal, thus increasing the spectral density at higher frequencies. For
a low loss channel, these higher frequency spectral components will be
evident at the receiver input and will therefore affect the receiver's
behavior. This is usually the case for a parallel interface (e.g.,
DDR2), for example. For a lossy channel, however, the loss increases
with frequency and so the channel attenuates the higher frequency
spectral components (including those generated by any nonlinearities)
more than it does the lower frequency spectral components (especially
those generated by the linear response).
In short, for high speed serial channels (i,e, those with high loss
channels), a linear driver model produces results which are more than
accurate enough to accurately predict link performance. Given this
assumption of linearity, the output impedance and output wave shape of a
driver are independent.
I do agree with you that the boundary between the algorithmic model and
the analog model is fuzzy; and in fact we have sometimes found it useful
to move that boundary for a given model. The only responsibility that
can be unequivocally assigned is that the analog model is solely
responsible for modeling the impedance presented to the interconnect
network. We generally assign the more complex behaviors such as
equalizer behavior or wave shape to the algorithmic model, however,
because the algorithmic model can easily, flexibly, and efficiently
implement behaviors that would be extremely difficult to implement in an
analog model.
Mike S.
On 10/14/2010 9:18 PM, Mike Steinberger wrote:
Scott-
Your conclusions are all correct. Thanks for the constructive insight.
Your follow-on questions bring out some important principles as well.
Consider that for either a driver or receiver, the combination of the
algorithmic and analog models should accomplish two goals:
1. The model voltage waveforms into or out of a matched load should
match the waveforms from the reference model (e.g., SPICE) under the
same conditions as closely as possible.
2. The impedance presented by the model to the interconnect network
should match the reference model's impedance under the same conditions
as closely as possible (i.e., need to present the correct reflection
coefficient to the interconnect network).
Thus, for a driver we make the analog model responsible for modeling the
output impedance and we do most of the wave shaping in the algorithmic
model. And yes, we do use recursive digital filters extensively in
driver algorithmic models for that purpose. The one exception is that we
typically try to set the output amplitude in the analog model so that it
is slightly more useful in an EDA tool that supports IBIS models but not
IBIS-AMI models.
This creates a bit of a fuzzy model boundary problem. Migrating all
analog wave shape control into the Algorithmic side of the model
guarantees that the IBIS Analog Model does not match the AMI model, and
the computed analog impulse response for the channel is wrong.
Reflection behavior for the Tx silicon will be incorrect, since
transmitter complex impedance is dependent upon not only the termination
resistance, but any pulse shaping that is performed in the termination
network. Since modern transmitters utilize peaking circuits like
T-coils between the distributed capacitance of the receiver and ESD
structures, you cannot move frequency response filtering of the output
waveform into the algorithmic model without seriously impacting the
broadband return loss of the transmitter. The best you can do is create
an approximate equivalent model that is good for a fixed line impedance.
Unfortunately, a package is anything but.
Lumping all wave shape control into the algorithmic section necessarily
violates your #2 goal.
You are also quite correct that IBIS-AMI modeling as it is currently
defined does not model common mode behavior. I submit that there is a
precedent in the sense that classic IBIS does such a poor job of
modeling differential mode transmission. This lack of direct modeling of
common mode behavior has two practical implications:
1. The model is really only defined for a single set of bias conditions
(e.g., common mode voltage). Therefore, well written documentation for a
model should state the bias conditions for which the model is valid, and
an intelligent user should make sure that the bias conditions in their
application match the bias conditions for the model.
2. Drivers can generate significant amounts of common mode noise, and
that noise can be coupled into adjacent differential paths. Similarly,
differential receivers have only a finite amount of common mode
rejection and a finite common mode rejection range.
We have some ideas for addressing common mode behavior and others have
made suggestions to the IBIS-ATM subcommittee as well. To date, however,
no system developer has expressed interest in these problems. If there
is someone who thinks common mode modeling is important enough to invest
in, especially if they're a system developer, we'd be happy to work with
them.
I'm a system developer. I express interest. If I'm interested, my
customers and our common customers are interested. Common mode modeling
will make or break 25 Gbps and higher signaling systems. You cannot
correctly see packaging common mode issues if the modes are not being
excited, and the modeling is not extended into the power domain of the
package. Above 5 GHz, packages begin to exhibit electromagnetic
non-locality problems.
Mike S.
On 10/14/2010 05:15 PM, Scott McMorrow for Jim Bell wrote:
Mike,
Thanks, I understand the tradeoff that you've made. Essentially what
you are saying is that anything that is buffered from the channel by a
high impedance may be considered for AMI DLL modeling. From my
perspective a table of s-parameters could just as easily have been used
to model the CTLE. But given that at the time there was no mechanism in
IBIS to include s-parameters, I can understand why it was placed in the
algorithmic section.
How about on the driver side? You did not respond to the wave shaping
portion of the question. What is your stance on analog wave shaping and
filtering that are part of the driver, and are not buffered from the
channel? It appears that you are incorporating these into the die
S-parameter models. Is that correct? Does this include any possible
asymmetric driver behavior?
Finally, on the driver and the receiver side, how are common mode
effects integrated into the channel modeling? I'm guessing that since
the receiver is buried in the AMI DLL that no degradation due to common
mode offset is modeled, since the information has been lost in the
differential transformation. However, common mode inducing skew and
waveform asymmetry could certainly be incorporated into the drive side
and propagated through the transmission channels.
best regards,
Scott
On 10/14/2010 5:19 PM, Mike Steinberger wrote:
Scott-
In a paper we gave at DesignCon2008 (attached), we described the
modeling of a CTLE in the algorithmic model of a receiver. In this
paper, we happened to refer to the filter as a "peaking filter" rather
than a "CTLE". This approach is equally applicable to a transmitter.
Since 2008, we have produced a number of IBIS-AMI receiver models for a
number of IP suppliers, and most of these receiver models contained a
CTLE. In each case, we have applied the techniques described in the
paper to the algorithmic model of the receiver, in each case, we have
achieved good correlation, and customers are using these models to do
real work.
I recommend against any attempt to incorporate the CTLE into the analog
model for two reasons:
1. The analog model will necessarily expose its internal structure
whereas the algorithmic model does not.
2. It's an awful lot easier to write these responses into the
algorithmic model than it is to design an analog circuit to produce the
desired response.
Have I addressed your question?
Thanks.
Mike Steinberger
On 10/14/2010 2:30 PM, Scott McMorrow wrote:
A question to the group.
Modern SERDES Tx and Rx circuits can employ both continuous time linear
equalizers (CTLE) at both the transmitter and receiver. In addition,
wave shape filtering is often employed to control the Tx output
spectrum. By definition, these circuits are in the analog domain, and
as a first approximation are linear time invariant (LTI).
How are these elements being handed with current AMI models within the
proposed comprehensive set of simulation flows, since IBIS Tx drivers
have only a partial ability to model wave shaping, and no ability to
model a CTLE? Are they currently being included as additional external
circuit models to be concatenated with the remainder of the Analog
channel?
Regards,
Scott
Other related posts:
