[ibis-macro] Clarifying why Deconvolution is required in a specific flow in BIRD 120
- From: "Walter Katz" <wkatz@xxxxxxxxxx>
- To: "'Morrison, Casey'" <cmorrison@xxxxxx>, "Arpad Muranyi" <Arpad_Muranyi@xxxxxxxxxx>
- Date: Fri, 19 Nov 2010 19:31:26 -0500 (EST)
Casey, Arpad,
De-convolution is required under BIRD 120 in only one specific case that I
will describe below:
The de-convolution case requires a channel with a Tx AMI model that has
Init_Returns_Impulse=True and GetWave_Exists=True. The Tx AMI_GetWave
function would include both the Tx equalization and affects such as
pattern dependent jitter.
The Rx AMI model has GetWave_Exists=False, and generates an output impulse
response which the IC vendor avers is sufficiently accurate to represents
the operation of the algorithmic part of the receiver. We must assume that
the Rx model will determine its equalization settings based on the impulse
response that is generated by Tx AMI_Init, as many Rx AMI_Init models do
today.
With this combination of Tx/Rx models, the EDA tool can run a statistical
analysis of the channel using Tx AMI_Init and Rx AMI_Init.
It is when users want the EDA tool to do a time domain simulation which
includes the pattern dependent jitter represented within the TX
AMI_GetWave function that de-convolution is required. If the output of Rx
AMI_Init was convolved with the output of Tx AMI_GetWave, the result would
double count the Tx Equalization since it is included in both the Tx
AMI_GetWave result, and the output of Rx AMI_Init (since it was included
in the input to Rx AMI_Init). One method to do a time domain simulation
that includes the accurate Tx AMI_GetWave functionality and does not
double count the Tx equalization is to determine the impulse response of
the Rx filter by using de-convolution between the input to Rx AMI_Init and
the output of Rx AMI_Init. EDA vendors have expressed, and have written
into BIRD 120 alternatives to doing de-convolution which range from "Do
not do time domain simulation - IT HURTS", to relying that the tool can
turn off optimization in Rx AMI_Init, and just run a solution space to
determine the optimum Rx tap settings, and then call Rx AMI_Init with the
optimal tap settings and an input that does not include the Tx
equalization.
Bottom line is either do not do time domain simulation, manually optimize
the Rx filter, or de-convolution.
There are alternatives, however. The model can implement the ability of
returning the impulse response of the filter alone, or the model developer
can implement an Rx AMI_GetWave, which would apply the Rx filtering
directly to the output fo Tx AMI_GetWave.
Walter
Other related posts:
