All,
Here are my take-aways from today’s discussion on reliable AMI flows
Flows are simplified if:
1. All AMI models are Dual models with the possible exception that terminal
Rx (and retimer Rx) can be Dual models or GetWave only models.
2. It is straightforward for the model maker to convert an Init Only model
to a Dual model. There was at least on objection to this claim.
3. Tx AMI_Init function should not adapt (i.e. their equalization is
independent of the input Impulse Response.
4. The Redriver flow is fundamentally flawed. The ultimate requirement is
that the Impulse Response input to the terminal Rx be the accumulated impulse
response of all upstream AMI models and channels. The current BIRD excludes all
contributions upstream of the last redriver Tx.
A BIRD is clearly required to correct the redriver flow.
A BIRD is required that states the the equalization applied by a Tx AMI_Init to
the Tx AMI_Init impulse response is independent of the contents of the Tx
AMI_Init impulse response. (I.e. the Tx AMI_Init does not adapt its
equalization based on the input impulse response).
In the last 12 years we have encountered only one Tx AMI_Init function that
tried to adapt it equalization, and the adaptation results caused over
equalization by the Tx that had to be accounted for by over equalization in the
Rx. Bottom line is that this model had to be run with its adaptation mode
turned off. This is what drove the back channel enhancements to IBIS.
I would replace the existing flows with simple flows that just daisy chain the
filters from the primary Tx AMI_Init and the terminal Rx AMI_Init. The
statistical back channel flow is identical, except the EDA tool needs to move
the BCI_parameter_out and BCI_parameter_in pointer. And the GetWave flow is
again almost identical with *wave replacing *impulse_matrix. All these flows
will assume all of the models are Dual models (except the terminal Rx and
retimer Rx) that can be GetWave only.
The BIRD can also say that using models that are Init Only or GetWave only
(except the terminal Rx or retimer Rx) is problematic and should be avoided.
Walter
Walter Katz
Work 508.647-7633
Cell 720.417-3762
[Description: Description: Visit MathWorks.com]