IBIS Macromodel Task Group Meeting date: 09 March 2021 Members (asterisk for those attending): Achronix Semiconductor Hansel Dsilva ANSYS: * Curtis Clark * Wei-hsing Huang Cadence Design Systems: * Ambrish Varma Ken Willis Jared James Google: Zhiping Yang Intel: Michael Mirmak Kinger Cai Alaeddin Aydiner Keysight Technologies: * Fangyi Rao * Radek Biernacki Ming Yan * Todd Bermensolo Rui Yang Luminous Computing * David Banas Marvell Steve Parker Mentor, A Siemens Business: * Arpad Muranyi Micron Technology: * Randy Wolff * Justin Butterfield Missouri S&T Chulsoon Hwang SAE ITC Jose Godoy SiSoft (Mathworks): * Walter Katz Mike LaBonte Teraspeed Labs: * Bob Ross Zuken USA: * Lance Wang The meeting was led by Arpad Muranyi. Curtis Clark took the minutes. -------------------------------------------------------------------------------- Opens: - None. ------------- Review of ARs: - Chulsoon to send his new PSIJ presentation to the ATM list. - Done. - Walter to send his Simpler Redriver Flow presentation to the ATM list. - Done. -------------------------- Call for patent disclosure: - None. ------------------------- Review of Meeting Minutes: Arpad asked for any comments or corrections to the minutes of the March 2nd meeting. Walter moved to approve the minutes. Ambrish seconded the motion. There were no objections. ------------- New Discussion: Redriver Flow Issues: Walter shared a new version of his "Simpler Redriver Flow" from the previous week. He noted the addition of several new slides. Walter said that if we are going to add a new Reserved Parameter and change the content of the impulse response matrix, we need to have a compelling reason. He said that if we can get away without doing so and find a solution that works with legacy models and supports the new or old flows, that would be ideal. Walter shared a stand-alone flow comparison slide (not part of the presentation emailed to ATM): Flow comparison slide - block diagrams of 3 Redriver flows - top - legacy flow - He said we all agree this is incorrect. - middle - a version of Walter's BIRD166 flow - bottom - the BIRD210 flow Walter said the assumption in this slide was that the final Rx is the only thing that adapts. Setting aside crosstalk for now, the middle and bottom flows generate the exact same answer as long as the Tx devices don't want to adapt based on the IRs of their downstream channels. Fangyi said one difference in the proposals, other than crosstalk support, was that BIRD210 preserves the legacy information in the IR matrix. Walter agreed that the bottom flow depicts BIRD210's new column information. Fangyi said one benefit of BIRD210 is that it provides all the legacy information, e.g., the Tx still gets its downstream channel's IR (legacy) as well as the cumulative upstream IR (a new column). Walter said he thought the legacy flow, where the Tx gets its downstream channel's IR, wasn't worth preserving. Ambrish said the problem is that we have Tx models that do adapt based on the downstream channel. Walter said that the only known example of a Tx model that tried to adapt based on its downstream channel IR didn't work very well at all. He said the original idea of passing the downstream channel's IR to the Tx was flawed. Ambrish and Fangyi said that choice should lie with the model maker. Walter said we may want to add some way for a Tx model to say it optimizes based on its downstream channel's IR. If we add that option, the tool could change its behavior accordingly. Walter said the blue dashed arrows in the slide's flow figures represent the LTI filter responses. He said, in the existing specification, columns 2 through N+1 are used to pass in N crosstalk IRs. The model than applies its LTI filter response to those columns. Nothing precludes the EDA tool from adding an extra crosstalk column and putting a unit IR (i.e., derivative of a 0 to 1 step function, a Dirac delta function - all values zero, except for one value that is equal to 1/(sample_interval)) in it. When this column is returned by the model, it contains the filter response. The EDA tool can then do all the convolutions it needs to do using the filter responses returned by the Init functions. There's no need for a new Reserved Parameter, and the tools could always have done this. Fangyi said everything Walter had described could be done with BIRD210. He said some Redriver models might optimize based on crosstalk, but Walter's proposal would not provide all of the crosstalk terms. Walter said the EDA tool could still pass in all the crosstalk terms and could add the extra unit impulse response column. Fangyi said it would corrupt the optimization if we passed in the unit IR as one of the crosstalk columns. The model would get the wrong answer, so we would need a way to make it understand the unit IR was not a true crosstalk IR. Walter said we could overcome that issue. He suggested that the model could detect that a unit IR was in one of the crosstalk columns and ignore that column for the purposes of optimization. Walter said handling all the crosstalk terms to compute the final contributions to the terminal Rx was tricky, especially if we consider NEXT as well as FEXT, and he wasn't sure even BIRD210 could capture them correctly. Fangyi said it could be proven with graph theory that BIRD210 allows all the contributions to be captured. The EDA tool will have all the information it needs to compute all of these terms with the BIRD210 flow. Fangyi said he'd reviewed the crosstalk combinations and defined the direct aggressor terms in his earlier presentations. Walter said it would be helpful to show, for one particular example, how the Init functions would be called to capture all the necessary crosstalk terms. David said it can be difficult to consider two competing proposals in their most general forms, and it would be helpful if we could come up with concrete examples to compare. The group did not arrive at a way to do this, as it could require tools and models supporting the proposed features. Ambrish asked if anyone had done any analysis on the cumulative effect we'd see from all of the various indirect crosstalk terms. Was it a first, second, third order effect? Also, what kind of effect do we see when models optimize based on crosstalk? Fangyi said the answer to the first question was highly dependent on the layout. Ambrish said the EDA tool can figure out the most important paths, and Fangyi agreed. Walter said the solid lines in his figures for direct FEXT aggressors could be as high as 10% and probably wouldn't be much higher. Therefore, the dotted lines representing the indirect terms were probably 10% of 10%, so about 1%. Arpad asked whether that "10% of 10%" was true if the channel layout were U shaped and went back toward the upstream channel. Walter said that would be NEXT, and that's really hard to deal with correctly. In response to Ambrish's second question about optimizing based on crosstalk, Fangyi said crosstalk optimization affects the gain at the receiver. You don't want to have too much gain and amplify the crosstalk too much. Ambrish said if you're doing crosstalk cancellation techniques you might adjust the gain. - Ambrish: Motion to adjourn. - Curtis: Second. - Arpad: Thank you all for joining. AR: Walter to send his latest Simpler Redriver Flow presentation to the ATM list. ------------- Next meeting: 16 March 2021 12:00pm PT ------------- IBIS Interconnect SPICE Wish List: 1) Simulator directives