[ibis-macro] Re: Question on dividing up the Tx behavior between the AMI and analog portions of the model

  • From: James Zhou <james.zhou@xxxxxxxxxx>
  • To: "Arpad_Muranyi@xxxxxxxxxx" <Arpad_Muranyi@xxxxxxxxxx>, 'IBIS-ATM' <ibis-macro@xxxxxxxxxxxxx>
  • Date: Wed, 14 Mar 2012 10:57:00 -0700

Hi Arpad,

I am sorry to hear that your "fingers are starting to hurt from all this typing 
:) ". I deeply appreciate your replies with attention to details and please 
don't feel obliged to do so in the future, especially when we also have other 
responsibilities at work place.

Sincerely,
James Zhou


From: ibis-macro-bounce@xxxxxxxxxxxxx [mailto:ibis-macro-bounce@xxxxxxxxxxxxx] 
On Behalf Of Muranyi, Arpad
Sent: Tuesday, March 13, 2012 6:37 PM
To: 'IBIS-ATM'
Subject: [ibis-macro] Re: Question on dividing up the Tx behavior between the 
AMI and analog portions of the model

James,

"Legacy IBIS models represented by [Ramp] or [Rising/Fall Waveform] are not LTI 
in general."

You hit the nail on its head, right on...  But not all of your
details are correct.

"suppose you drive  an IBIS model with 1V pulse, and you get 1V output.  
However if you drive it with 0.1V pulse, instead of getting 0.1V output, you 
get 0V"

This is incorrect.  Let's separate this into two parts.  The logic
level of '1' or '0' in the digital stimulus has no thresholds, it is
either logic '1' or '0'.  There is no voltage or amplitude associated
with this.  This stimulus indicates the t=0 for the transition the
[Ramp] and waveform based buffer model does.  The model's amplitude
is determined by the reference voltages of the I-V tables and the
load impedance.  There are no ifs and buts as far as this stimulus,
it either switches or not.

If by "suppose you drive  an IBIS model with 1V pulse" you are referring to the
D_to_A converter's voltage levels which is used to drive an analog
circuit block representing the buffer's impedance, there are still no
thresholds associated with it.  The analog circuit (or S-parameter
model) that follows this converter will generate a proportional output
voltage, no matter what the amplitude of the D_to_A converter is.  So
there is no such thing that a 0.1V signal will have 0 output.

"By definition, [Ramp] or [Rising/Falling Waveform] are only triggered by an 
input event at time=t0. All input amplitude variations have no impact on the 
output (unless they cause a trigger of Rising/Falling waveforms). Such a system 
cannot be represented by an impulse response. "

Again, there are no thresholds, so the above sentence is not quite correct.

But think about it this way:  The system's transfer function doesn't involve
the logic '1' or '0' of the stimulus (or the transition between these states)
because these logic states are not directly related to the voltages or currents
in the system.  The transfer function of an analog driver (or even receiver)
consists of its I-V relationship (and time) between the signal pad and the
supply rails.  If you force the voltage on the signal pad, the buffer will
respond with a current, or vice versa.  The response you get may depend on
the logic state of the digital input stimulus and the bit pattern that
changes it, which is a variation with respect to time.  So in short, the
transfer function of a buffer is pretty much impedance as a function of
voltage and time.  (You may also add frequency to that, but that really
doesn't change the above discussion much).

Some sources of non-linearities may involve a nonlinear I-V relationship,
which looks like a curve instead of straight line on the I-V plot, and
time variances may involve a changing impedance with respect to time,
and similar artifacts.  These effects are assumed to be negligible in
the current IBIS-AMI specification whether it is true or not.  Even though
legacy analog models (using I-V curves) have the capabilities to describe
a non-linear I-V relationship, for AMI purposes it is advisable not to
make such non-linear analog models.  Also, even though legacy IBIS models
do have the capability to describe asymmetric rise and fall times, it
is better not to make models exhibiting such behaviors.

The new analog modeling proposals which make use of the D_to_A converters
as ideal sources and S-parameters or similar linear circuits for modeling
the impedance and rise/fall time of the buffer do away with the possibility
to make non LTI analog buffer models, but these modeling techniques are not
going to do away with the facts of real life, which is generally non-LTI.

In order to include such capabilities in AMI-style algorithmic simulations,
we would have to revisit the fundamentals of our AMI approach.  My colleague,
Vladimir gave a presentation on that topic a while ago, but so far the
ideas presented there didn't catch on.

http://www.vhdl.org/pub/ibis/macromodel_wip/archive/20091014/vladimirdmitrievzdorov/How%20to%20account%20for%20non-LTI%20of%20Tx%20analog%20buffer%20in%20IBIS%20AMI%20flow/AMI_Tx_With_Admittance.pdf

The question on how anyone goes about extracting the impulse response hAC(t)
with or without the [Ramp] or waveforms and the I-V curves, or employing any
other technique is a very complicated topic.  It is best left to the rocket
scientists of each EDA vendor.  Who knows, they may even have some snake oil
in their secret sauce.  As long as each vendor can claim that they have the
best and most accurate algorithms to generate impulse responses, their
customers will be happy.  The IBIS specification is only here to make sure
that given the same impulse response the same model should pretty much get
you the same results, no matter on whose simulator the model is executed.
The question of how good the impulse response is falls into the category
of garbage in - garbage out discussions as far as the AMI spec is concerned.

I hope this helps...

Arpad
P. S.   Please try not to ask more questions unless absolutely necessary,
my fingers are starting to hurt from all this typing ( :) ).
===============================================================================

From: James Zhou [mailto:james.zhou@xxxxxxxxxx]
Sent: Tuesday, March 13, 2012 6:34 PM
To: Muranyi, Arpad; 'IBIS-ATM'
Subject: RE: [ibis-macro] Re: Question on dividing up the Tx behavior between 
the AMI and analog portions of the model

Arpad,

One more comment on  " to generate hAC(t)... the EDA tool would give a stimulus 
to the Tx analog model corresponding to a rising or falling step function, such 
as a digital '0' and '1'  or  '1' and '0' to get a rising or falling step 
waveform"

I think there are some potential problems to obtain hAC(t) in this way. First 
of all, hAC(t) can only represent LTI block.  Legacy IBIS models represented by 
[Ramp] or [Rising/Fall Waveform] are not LTI in general. One simple example is 
shown here:  suppose you drive  an IBIS model with 1V pulse, and you get 1V 
output.  However if you drive it with 0.1V pulse, instead of getting 0.1V 
output, you get 0V.

If the above process is permissible by the Specification to generate hAC(t) 
from [Ramp] or [Rising/Falling Waveform], some linearization process must be 
used either implicitly or explicitly.  This linearization process may not be 
trivial for any arbitrary legacy IBIS model and I think the Specification 
should provide sufficient information such that all EDA tools should end up 
with the same hAC(t) from the same data.

By definition, [Ramp] or [Rising/Falling Waveform] are only triggered by an 
input event at time=t0. All input amplitude variations have no impact on the 
output (unless they cause a trigger of Rising/Falling waveforms). Such a system 
cannot be represented by an impulse response.

Regards,
James Zhou


From: ibis-macro-bounce@xxxxxxxxxxxxx<mailto:ibis-macro-bounce@xxxxxxxxxxxxx> 
[mailto:ibis-macro-bounce@xxxxxxxxxxxxx]<mailto:[mailto:ibis-macro-bounce@xxxxxxxxxxxxx]>
 On Behalf Of Muranyi, Arpad
Sent: Tuesday, March 13, 2012 2:02 PM
To: 'IBIS-ATM'
Subject: [ibis-macro] Re: Question on dividing up the Tx behavior between the 
AMI and analog portions of the model

James,

Regarding:

"please clarify what is actually driving  the input terminals of Tx analog 
block (which is the first stage of the entire analog channel, represented by 
hAC(t))",

there are several different ways to generate hAC(t), but one way to
do it is to run a time domain simulation using a step function and
take the derivative of the resulting waveform at the Rx.  In order
to do this, the EDA tool would give a stimulus to the Tx analog model
corresponding to a rising or falling step function, such as a digital
'0' and '1'  or  '1' and '0' to get a rising or falling step waveform.

How this is implemented is entirely up to the EDA vendor, that's why
the IBIS specification doesn't give you these details.  But the point
is that this stimulus has nothing to do with x(t) which is what goes
into AMI_GetWave during an AMI simulation.


Regarding:

"(a)  Instead of  "driving the channel topology by the Tx I-V, [Ramp], or V-t 
based analog model", the "digital stimulus" actually drives the Tx AMI input. 
The Tx AMI output should drive these analog models, if such practices are 
allowed by the Specification. It neither makes sense to me (to drive a legacy 
IBIS model with x(t) or Tx AMI output) nor specified by IBIS 5.0 or BIRDs that 
this is permissible and doable."

Which digital stimulus are you talking about in "the "digital stimulus" 
actually drives the Tx AMI input"?
There are two digital stimuli.  One that is used to stimulate the
analog Tx model to generate the impulse response (hAC(t)) for the
channel, and another one that is the input to the Tx AMI_GetWave
function during the AMI simulation.

Why do you say "The Tx AMI output should drive these analog models"?  I don't
see that in the AMI flow diagram.  Even though you might want to do
that in your approach, this is not what the IBIS AMI flow suggests.


Regarding: "since the "digital stimulus" x(t) is in fact amplitude continuous, 
if it is used to drive the D_to_A in Tx",
this statement is not correct.  x(t) is not used to drive the D_to_A
(or the input) of the Tx analog model.  It is used as the input to the
Tx GetWave function.

I hope this helps.

Arpad
=======================================================================


From: James Zhou 
[mailto:james.zhou@xxxxxxxxxx]<mailto:[mailto:james.zhou@xxxxxxxxxx]>
Sent: Tuesday, March 13, 2012 1:15 PM
To: Muranyi, Arpad; 'IBIS-ATM'
Subject: RE: [ibis-macro] Re: Question on dividing up the Tx behavior between 
the AMI and analog portions of the model

Hi Arpad,

I agree with the first part of your response, which basically states that x(t) 
= b(t) * p(t) and, x(t) drives Tx AMI block. In case of using AMI_GetWave, x(t) 
is passed to AMI_GetWave through *wave variable. Even though labeled as 
"digital stimulus", x(t) is amplitude continuous and time discrete.

With reference to "The analog model was used in generating the channel's impulse
response hAC(t), but it is not driven by the output of Tx
GetWave as you concluded",
If this is the case, please clarify what is actually driving  the input 
terminals of Tx analog block (which is the first stage of the entire analog 
channel, represented by hAC(t)).

With reference to "As far as I know the analog models are
exercised the same way as in a usual legacy IBIS simulation using the
tool's digital stimulus (though the D_to_A converters in Tx), driving
the channel topology by the Tx I-V, [Ramp] or V-t based analog model,
to obtain an impulse response as seen at the Rx",
there are several issues here:
(a)  Instead of  "driving the channel topology by the Tx I-V, [Ramp], or V-t 
based analog model", the "digital stimulus" actually drives the Tx AMI input. 
The Tx AMI output should drive these analog models, if such practices are 
allowed by the Specification. It neither makes sense to me (to drive a legacy 
IBIS model with x(t) or Tx AMI output) nor specified by IBIS 5.0 or BIRDs that 
this is permissible and doable.
(b)  since the "digital stimulus" x(t) is in fact amplitude continuous, if it 
is used to drive the D_to_A in Tx, all amplitude information would be lost, 
except the rise/fall transitions.

Thanks,
James Zhou



From: ibis-macro-bounce@xxxxxxxxxxxxx<mailto:ibis-macro-bounce@xxxxxxxxxxxxx> 
[mailto:ibis-macro-bounce@xxxxxxxxxxxxx]<mailto:[mailto:ibis-macro-bounce@xxxxxxxxxxxxx]>
 On Behalf Of Muranyi, Arpad
Sent: Monday, March 12, 2012 9:36 PM
To: 'IBIS-ATM'
Subject: [ibis-macro] Re: Question on dividing up the Tx behavior between the 
AMI and analog portions of the model

James,

Knowing that Walter is enjoying the IEEE meetings in Hawaii,
I will try to help a little.  Please refer to the flow
diagram that was used in the discussions towards the new
flow.  See pg. 17 in this presentation:

http://www.vhdl.org/pub/ibis/macromodel_wip/archive/20100713/toddwesterhoff/IBIS-AMI%20Flows/Flows_July2010-v2.pdf

Note that the input to the Tx GetWave function is labeled
"Digital stimulus" and x(t).  Also note that x(t) is
explained in more detail in this presentation on pg. 13-14:

http://www.vhdl.org/pub/ibis/macromodel_wip/archive/20061212/toddwesterhoff/Serial%20Link%20Terminology/serial_link_terminology.pdf

where "TX data" is b(t) convolved with p(t) which we later
simply referred to as x(t) in our flow diagrams.

Please note also that on pg. 17 of the first presentation
above, the output of Tx GetWave is convolved with the channel
impulse response hAC(t) and the result of that is the input
to Rx GetWave.  There is no analog model present in this flow.
The analog model was used in generating the channel's impulse
response hAC(t), but it is not driven by the output of Tx
GetWave as you concluded.

So your conclusions that:

" Equivalently, the legacy IBIS [Ramp] and [Rising/Falling Waveform] keywords 
have no roles in this process."

and

"D_to_A  is not needed in IBIS AMI, or at least in between Tx AMI block and its 
analog block. Because the output of Tx AMi is already analog, why bother using 
a D_to_A? "

seem to be incorrect to me.  As far as I know the analog models are
exercised the same way as in a usual legacy IBIS simulation using the
tool's digital stimulus (though the D_to_A converters in Tx), driving
the channel topology by the Tx I-V, [Ramp] or V-t based analog model,
to obtain an impulse response as seen at the Rx.

Of course there are other techniques to achieve equivalent results, but
it is not the role of the IBIS specification to enumerate all possible
methods in the trade.

I hope this helps...

Arpad
=========================================================================



From: ibis-macro-bounce@xxxxxxxxxxxxx<mailto:ibis-macro-bounce@xxxxxxxxxxxxx> 
[mailto:ibis-macro-bounce@xxxxxxxxxxxxx]<mailto:[mailto:ibis-macro-bounce@xxxxxxxxxxxxx]>
 On Behalf Of James Zhou
Sent: Monday, March 12, 2012 9:02 PM
To: Walter Katz; Terry.Chen@xxxxxxxxxx<mailto:Terry.Chen@xxxxxxxxxx>; 
DBanas@xxxxxxxxxx<mailto:DBanas@xxxxxxxxxx>; 'IBIS-ATM'
Subject: [ibis-macro] Re: Question on dividing up the Tx behavior between the 
AMI and analog portions of the model

Hi Walter,

Thanks for your response and clarification. The issues can be separated as the 
following:

(1) "determine the Impulse Response of a channel": I fully agree with you that 
"Tx analog portion takes the output of the algorithmic section of an AMI 
model". Equivalently, the legacy IBIS [Ramp] and [Rising/Falling Waveform] 
keywords have no roles in this process. This is a subject that has caused much 
confusion in the end-user community and we need to clarify it in terms of both 
(a) what are the intentions of model makers when IBIS AMI analog model data is 
put in [Ramp] [Raising/Falling Waveform] keywords and (b) What should the EDA 
tools do about those data in IBIS AMI modeling (i.e. when deriving impulse 
response)

(2)  "the fundamental confusion of using [External Model] with IBIS-ISS 
subckts": in my opinion,  D_to_A  is not needed in IBIS AMI, or at least in 
between Tx AMI block and its analog block. Because the output of Tx AMi is 
already analog, why bother using a D_to_A?  ( I understand D_to_A is useful for 
general IBIS [External Model] not involving AMI).

(3)  "one must be very careful to make sure that return loss is properly 
accounted for": I fully agree. IBIS Specification should provide clear 
guidelines on how to achieve this both at model creation time and model 
simulation time. I think enforcing reverse isolation is too restrictive for 
existing silicon implementation and it will cause larger errors when return 
loss is lower. It is much simpler and better to require the disclosure 
(knowledge) of Tx AMI block output impedance (which is assumed to be high 
impedance by many, anyway) and, not imposing any artificial requirements on the 
AMI analog block (such as reverse isolation which in fact does not exist in 
many silicon).

Thanks,
James Zhou


From: Walter Katz [mailto:wkatz@xxxxxxxxxx]<mailto:[mailto:wkatz@xxxxxxxxxx]>
Sent: Monday, March 12, 2012 4:50 PM
To: James Zhou; Terry.Chen@xxxxxxxxxx<mailto:Terry.Chen@xxxxxxxxxx>; 
DBanas@xxxxxxxxxx<mailto:DBanas@xxxxxxxxxx>; 'IBIS-ATM'
Subject: RE: [ibis-macro] Re: Question on dividing up the Tx behavior between 
the AMI and analog portions of the model

James,

The IBIS analog model is only useful to determine the Impulse Response of a 
channel. You are absolutely correct that in reality a SerDes Tx analog portion 
takes the output of the algorithmic section of an AMI model to drive an analog 
model (e.g. an on-die S-parameter s4p, a simpler RC circuit at described in 
BIRD 122, or an ISS subckt as defined in BIRD 116). IBIS AMI assumes an LTI 
channel, and IBIS-ISS defines all of the LTI elements available in HSPICE.

This is the fundamental confusion of using [External Model] with IBIS-ISS 
subckts describing the analog section. As written now, BIRD 116 identifies the 
input of the Tx ISS subckt using the D_to_A statement, which essentially 
defines a voltage swing and rise time - equivalent to Ramp. The correct 
interpretation is that the D_to_A statement is only to define the input to the 
Tx analog circuit, and is valid to determine the Impulse Response of the 
channel.

Because of silicon drivers are in fact LTI, one can do the shaping of the 
waveform in the algorithmic section, but one must be very careful to make sure 
that return loss is properly accounted for.

Walter

From: ibis-macro-bounce@xxxxxxxxxxxxx<mailto:ibis-macro-bounce@xxxxxxxxxxxxx> 
[mailto:ibis-macro-bounce@xxxxxxxxxxxxx]<mailto:[mailto:ibis-macro-bounce@xxxxxxxxxxxxx]>
 On Behalf Of James Zhou
Sent: Monday, March 12, 2012 5:11 PM
To: Terry.Chen@xxxxxxxxxx<mailto:Terry.Chen@xxxxxxxxxx>; 
DBanas@xxxxxxxxxx<mailto:DBanas@xxxxxxxxxx>; 'IBIS-ATM'
Subject: [ibis-macro] Re: Question on dividing up the Tx behavior between the 
AMI and analog portions of the model

Hi David and Terry,

Both of your emails mentioned "analog IBIS model" and "IBIS-analog portion" 
represented by [Ramp] and/or [Rising/Falling Waveform] keywords in IBIS file.  
However, these "analog" IBIS models only take digital input signals,  as stated 
in IBIS Specification 5.0, page 71-72 and section 6b. The output of the 
"analog" IBIS model is not capable of tracking the amplitude changes in the 
input (other than a rise/fall transition). It would not make sense to feed the 
Tx AMI output to such digital inputs based on IBIS Specification 5.0.

If this approach of using [Ramp] and/or [Rising/falling Waveform] keywords to 
represent "analog IBIS model" is adopted by IBIS AMI flow, some clarification 
is needed on how to interpret and implement it.

Regards,
James Zhou
QLogic Corp.


From: ibis-macro-bounce@xxxxxxxxxxxxx<mailto:ibis-macro-bounce@xxxxxxxxxxxxx> 
[mailto:ibis-macro-bounce@xxxxxxxxxxxxx]<mailto:[mailto:ibis-macro-bounce@xxxxxxxxxxxxx]>
 On Behalf Of Chen, Terry
Sent: Thursday, March 08, 2012 10:33 AM
To: DBanas@xxxxxxxxxx<mailto:DBanas@xxxxxxxxxx>; 'IBIS-ATM'
Subject: [ibis-macro] Re: Question on dividing up the Tx behavior between the 
AMI and analog portions of the model

Hi David,

Actually I am interested in other's response to this question as well...

But, for the TX Driver I am currently modeling, I am doing exactly what you 
have prescribed and using the IBIS-analog portion as effectively an ideal step 
function (by setting my ramp with extremely high rise/fall dv/dt) and letting 
the step response filter inside my AMI model to shape my output waveform. Now, 
I am not sure if this is the "right" or "ideal" way to do it, but I am getting 
a reasonably good correlation in my Re-driver model with the actual lab 
measurements (the max jitter mismatch is < 8ps).

I hope this is at least an useful data point for you.

Regards,
Terry

From: ibis-macro-bounce@xxxxxxxxxxxxx<mailto:ibis-macro-bounce@xxxxxxxxxxxxx> 
[mailto:ibis-macro-bounce@xxxxxxxxxxxxx]<mailto:[mailto:ibis-macro-bounce@xxxxxxxxxxxxx]>
 On Behalf Of David Banas
Sent: Thursday, March 08, 2012 1:15 PM
To: 'IBIS-ATM'
Subject: [ibis-macro] Question on dividing up the Tx behavior between the AMI 
and analog portions of the model

Hi all,

Is it customary to split up the Tx behavior, such that the FFE is modeled in 
the AMI model and the pulse shaper in the analog model?
Or, is there a different dividing line that has been identified as "best 
practice".
(Or, am I completely off in the weeds?)

The context for this question: I just managed to get good correlation between 
our latest Tx AMI model and the HSPICE model.
And then I realized that, having dumped all of the behavior into the AMI model, 
I would need to put an ideal step function into the V-T curves of the analog 
IBIS model. And I wasn't sure that would be a good idea. (I'm guessing that 
that would reek havoc in most simulators; is that correct?)

Thanks,

David Banas
Sr. Member Technical Staff
Altera<http://www.altera.com/>
+1-408-544-7667 - desk

Did you know Altera offers over 150 free online technical training 
courses<http://www.altera.com/servlets/searchcourse?coursetype=Online&WT.mc_id=t9_ot_mi_mi_tx_a_311>?
 Take one today!


________________________________
Confidentiality Notice.
This message may contain information that is confidential or otherwise 
protected from disclosure. If you are not the intended recipient, you are 
hereby notified that any use, disclosure, dissemination, distribution, or 
copying of this message, or any attachments, is strictly prohibited. If you 
have received this message in error, please advise the sender by reply e-mail, 
and delete the message and any attachments. Thank you.

________________________________
This message and any attached documents contain information from QLogic 
Corporation or its wholly-owned subsidiaries that may be confidential. If you 
are not the intended recipient, you may not read, copy, distribute, or use this 
information. If you have received this transmission in error, please notify the 
sender immediately by reply e-mail and then delete this message.

________________________________
This message and any attached documents contain information from QLogic 
Corporation or its wholly-owned subsidiaries that may be confidential. If you 
are not the intended recipient, you may not read, copy, distribute, or use this 
information. If you have received this transmission in error, please notify the 
sender immediately by reply e-mail and then delete this message.

________________________________
This message and any attached documents contain information from QLogic 
Corporation or its wholly-owned subsidiaries that may be confidential. If you 
are not the intended recipient, you may not read, copy, distribute, or use this 
information. If you have received this transmission in error, please notify the 
sender immediately by reply e-mail and then delete this message.

________________________________
This message and any attached documents contain information from QLogic 
Corporation or its wholly-owned subsidiaries that may be confidential. If you 
are not the intended recipient, you may not read, copy, distribute, or use this 
information. If you have received this transmission in error, please notify the 
sender immediately by reply e-mail and then delete this message.

________________________________
This message and any attached documents contain information from QLogic 
Corporation or its wholly-owned subsidiaries that may be confidential. If you 
are not the intended recipient, you may not read, copy, distribute, or use this 
information. If you have received this transmission in error, please notify the 
sender immediately by reply e-mail and then delete this message.

Other related posts:

  1. Home
  2. ibis-macro
  3. Archive
  4. 03-2012