[ibis-macro] Re: An AMI Overview
- From: "Mellitz, Richard" <richard.mellitz@xxxxxxxxx>
- To: "dkirsanov@xxxxxxxxxx" <dkirsanov@xxxxxxxxxx>, "ibis-macro@xxxxxxxxxxxxx" <ibis-macro@xxxxxxxxxxxxx>
- Date: Tue, 20 Oct 2009 10:56:48 -0600
I completely agree with Danil. For the reasons stated below, my experience
suggests de-convolution often yields results that are drawn with a spray can.
It's not bad for painting pictures, but more often than not reuse of the
de-convolved data takes lots of care.
... Rich Mellitz, Intel
From: ibis-macro-bounce@xxxxxxxxxxxxx [mailto:ibis-macro-bounce@xxxxxxxxxxxxx]
On Behalf Of Danil Kirsanov
Sent: Tuesday, October 13, 2009 11:01 AM
To: ibis-macro@xxxxxxxxxxxxx
Subject: [ibis-macro] Re: An AMI Overview
Quick comment on de-convolution.
We have some experience with it in our other projects, and it is not as
straightforward as it looks. Basically, de-convolution is an ill-posed problem
that is very sensitive to the numerical noise and particular impulse response
used. The theoretical approach of dividing the fourier spectrums does not work
very well in practice, resulting in very large error, and require sophisticated
regularization that is not guaranteed to work in all cases (see Wiener filter
as a simple example).
I believe this approach should be avoided if possible, since in many cases it
will be extremely hard to produce a reliable answer.
Best,
Danil
From: ibis-macro-bounce@xxxxxxxxxxxxx [mailto:ibis-macro-bounce@xxxxxxxxxxxxx]
On Behalf Of Walter Katz
Sent: Monday, October 12, 2009 7:39 PM
To: fangyi_rao@xxxxxxxxxxx; ibis-macro@xxxxxxxxxxxxx
Subject: [ibis-macro] Re: An AMI Overview
Fangyi,
By deconvolution ( http://en.wikipedia.org/wiki/Deconvolution ).
The input to Rx Init is h_AC*h_TEI, the output of Rx Init is h_AC*h_TEI*h_REI.
Deconvolution takes the Fourier Transform of the input and output impulse
responses, then divides the coefficients of the two Fourier Transforms, and
then does an inverse Fourier Transform to get h_REI. We do this for this flow
because we have no other choice. That is why I am proposing to add to Rx Init
models the ability to just return h_REI.
Walter
Walter Katz
303.449-2308
Mobile 720.333-1107
wkatz@xxxxxxxxxx
www.sisoft.com
-----Original Message-----
From: fangyi_rao@xxxxxxxxxxx [mailto:fangyi_rao@xxxxxxxxxxx]
Sent: Monday, October 12, 2009 6:55 PM
To: wkatz@xxxxxxxxxx; ibis-macro@xxxxxxxxxxxxx
Subject: RE: [ibis-macro] Re: An AMI Overview
Hi, Walter;
Thanks for your explanation. Regarding point 2, I still think if model vendors
are allowed to modify impulse any way they want, then following scenario will
be broken.
Tx has GetWave.
Tx Init modifies h_AC and returns h_AC_Tx=h_AC*h_TEI
Rx does NOT have GetWave.
h_AC_Tx is passed into Rx Init.
Rx Init modified h_AC_Tx and returns h_AC_Tx_Rx= h_AC*h_TEI + h_REI
Tx GetWave output is to be convolved with h_AC_Rx=h_AC + h_REI, which is the
combined impulse of channel and Rx.
Without knowing how Rx Init modifies the impulse, how can EDA tools remove the
h_TEI portion from h_AC_Tx_Rx to get h_AC_Rx?
Thanks,
Fangyi
From: Walter Katz [mailto:wkatz@xxxxxxxxxx]
Sent: Monday, October 12, 2009 3:04 AM
To: RAO,FANGYI (A-USA,ex1); ibis-macro@xxxxxxxxxxxxx
Subject: RE: [ibis-macro] Re: An AMI Overview
Fangyi,
1. Even if a model is non-LTI, it may have an LTI approximation. If the
Init call returns an LTI approximation, then the EDA tool can use "Statistical"
methods to do very fast analysis, realizing that the GetWave time domain
simulation will still need to be made to confirm the result. If the Init call
does not return an LTI approximation, then the EDA tool must use time domain
methods only to analyze the channel.
2. The current IBIS 5.0 specification has words like (section 2.1.6):
| 6. AMI_Init parses the configuration parameters, allocates dynamic
| memory, places the address of the start of the dynamic memory in
| the memory handle, computes the impulse response of the block and
| passes the modified impulse response to the EDA tool. The new
| impulse response is expected to represent the filtered response.
Since returning an impulse response assumes LTI (or an LTI approximation), the
model maker can do this calculation any way he chooses. One way to do this
calculation convolution. We spent much time determining if we should use the
word convolution, concatenation, combination, and ended up with the word
modified. It would probably be worth adding a paragraph to the document that
explicitly says that whenever convolution is used, it is simply to represent
the result of the function to be preformed, and the model maker or EDA vendor
can use any mathematical method deemed appropriate for performance, accuracy or
other algorithmic reasons.
Walter
Walter Katz
303.449-2308
Mobile 720.333-1107
wkatz@xxxxxxxxxx
www.sisoft.com
-----Original Message-----
From: ibis-macro-bounce@xxxxxxxxxxxxx
[mailto:ibis-macro-bounce@xxxxxxxxxxxxx]On Behalf Of fangyi_rao@xxxxxxxxxxx
Sent: Sunday, October 11, 2009 8:22 PM
To: wkatz@xxxxxxxxxx; ibis-macro@xxxxxxxxxxxxx
Subject: [ibis-macro] Re: An AMI Overview
Hi, Walter;
Thanks for your clarification of AMI methodology. I'd like to see the overview
or the spec also states clearly about following two topics.
1. For non-LTI models that have GetWave, how does AMI_Init enable model
developers to also support statistical simulation using LTI approximation?
2. Do we explicitly assume the modified impulse returned by AMI_Init is
the CONVOLUTION between input impulse and the equalizer? I know at least one
model vendor would not be happy about this assumption because their modified
impulse is not from convolution.
Regards,
Fangyi
From: ibis-macro-bounce@xxxxxxxxxxxxx [mailto:ibis-macro-bounce@xxxxxxxxxxxxx]
On Behalf Of Walter Katz
Sent: Saturday, October 10, 2009 2:20 PM
To: IBIS-ATM
Subject: [ibis-macro] An AMI Overview
All,
We have all been getting into the nuts and bolts of AMI. I think we need to
step back and understand what AMI is all about. The following has been said in
many ways in many places, but I think it is appropriate to review it at this
time. I do not know if this belongs in the IBIS AMI specification, but it sure
would be nice.
Walter
AMI Overview
The analysis of high speed (>4 Gbps) offers some simplifications and some
challenges. The simplifications arise from the fact the driver (Tx) final stage
output, and the receiver (Rx) input along with the interconnect between them
can be treated as Linear and Time Invariant (LTI) "Analog-Channel". This allows
various mathematical techniques to be applied improving simulation performance
by orders of magnitude. The complication that has led to the development of the
AMI standard arises from the need for complex signal processing in both the Tx
before the final stage driver and particularly the Rx after the Rx receiver.
The good news is that the signal processing does not interact with the
Tx/Interconnect/Rx Analog-Channel. The signal processing that is going on is
considered important Intellectual Property (IP) of the IC vendors, and this
along with the need for simulating millions of bits and the IP considerations
led to the AMI standard.
A SerDes-Channel consists of Tx signal processing, Analog-Channel and Rx signal
processing. The Tx AMI model represents the Tx signal processing, an impulse
response represents the Analog-Channel, and the Rx AMI model represents the Rx
signal processing.
The Analog-Channel is represented as an impulse response hAC(t). It is the
differential mode impulse response of the interconnects between the Tx and the
Rx and includes the reactive load (impedance) of the Tx final stage driver and
the Rx receiver. This impulse response can be created using many mathematical
techniques, including, but not limited to simulation and TDR measurement.
Traditional IBIS does not model differential LTI buffers, this will be
addressed when I introduce new AMI reserved parameters. This is also related to
the topic of Vladimir's presentation this Tuesday.
The Tx AMI model and Rx AMI model may themselves be either LTI or non-LTI. If
they are LTI they can be represented accurately by an impulse response. If they
are not LTI, they can be approximated by an impulse response. AMI models are
delivered as executable code in the form of a Shared Object (SO) or a
Dynamically Linked Library (DLL), and an .ami ASCII file. All AMI DLL's have
an AMI_Init entry, and an AMI_Close entry. This is all that is required if the
model is LTI. If a model is non-LTI then is must also have an AMI_GetWave. If a
model does have an AMI_GetWave then the model make is telling the EDA tool that
the model is non-LTI and that using the just impulse response from the AMI_Init
is not deemed sufficient to accurately model the channel.
The .ami file tells the EDA tool if the model has an AMI_GetWave entry, and
sufficient information to configure the model, and to analyze the results that
the model generates.
When a model is LTI it does not need a GetWave entry. When LTI, the only
information that need be abstracted from the model is the impulse response of
the models equalization. The input to the Tx AMI_Init function is the impulse
response of the channel. The input to the Rx AMI_Init function is the impulse
response of the channel combined with the impulse response of the Tx
equalization obtained from the call to Tx AMI_Init.
When a model is non-LTI (or more importantly, the IC Vendor believes that the
non-LTI behavior is important) then it does require an AMI_GetWave entry. The
AMI_Init entry is used to return to the EDA tool an approximate impulse
response of the models equalization, and to pass on initialization information
to be used by the AMI_GetWave entry. The AMI_GetWave entry is then called
repeatedly with sequential blocks of waveforms. The waveforms are that are
input to Tx AMI_GetWave indicate the transition times of the digital stimulus
input to the Tx equalization circuitry. The waveforms that are output of Tx
AMI_GetWave are the waveforms that drive the Analog-Channel. The waveform that
is input to Rx AMI_GetWave is the waveform that is the input to the Rx buffer.
The output of the Rx AMI_GetWave is the waveform at the Rx decision point, and
optionally clock ticks indicating the location of each recovered clock. The EDA
tool processes the waveform at the Rx decision point, and either uses the clock
ticks or the Clock Recovery Mean and Rj to generate bathtub curves, BER and
other channel data.
A fundamental principle of AMI modeling is that every EDA platform (both
software and hardware) will give the same results when presented with the same
Analog-Channel impulse response, the same AMI model conditions, and the same
input stimulus pattern. Each EDA platform may differ on how its sets the Tx and
Rx AMI model conditions, the stimulus pattern, how it creates the
Analog-Channel impulse response, and how it processes the resulting outputs.
Walter Katz
303.449-2308
Mobile 720.333-1107
wkatz@xxxxxxxxxx
www.sisoft.com
Other related posts:
