IBIS Macromodel Task Group Meeting date: 27 April 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 Micron Technology: * Randy Wolff * Justin Butterfield Missouri S&T Chulsoon Hwang Siemens EDA (Mentor): * Arpad Muranyi 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: - Walter to modify the new Redriver proposal as discussed during the meeting and submit it as BIRD211.1. - Done. It was submitted to the Open Forum and introduced at the April 23rd Open Forum meeting. -------------------------- Call for patent disclosure: - None. ------------------------- Review of Meeting Minutes: Arpad asked for any comments or corrections to the minutes of the April 20th meeting. Randy moved to approve the minutes. Walter seconded the motion. There were no objections. ------------- New Discussion: BIRD211.1: Arpad noted that the Open Forum had approved a motion to schedule a vote on this BIRD at its next meeting on May 14th. He said he hoped we could finish discussion on any cleanup of the BIRD and get it submitted by Friday, April 30th, which is the deadline if we want to vote on it at the next Open Forum meeting. Walter shared a new BIRD211.2_draft1 he had prepared. He had sent it to ATM the day before, and it contains corrections for typos and errors found in BIRD211.1. Walter shared the BIRD draft and focused on the block diagram depicting the three proposed flows. Walter noted that Radek had said during the Open Forum discussion that we might still have to review the need for flow B (the "Combined" flow). Arpad and Curtis said Radek's comments had surprised them, and they thought we had agreed to keep flow B at the last ATM meeting. Fangyi said he hadn't attended the Open Forum meeting and hadn't spoken with Radek since. So, he couldn't comment on Radek's statement. Walter noted that the first flow, which is invoked if the parameter is set to "IBIS7.0" is a corrected version of the IBIS 7.0 flow. The output of Rx1 Init is combined with the output of Tx2 Init, and the result is passed to Rx2's Init. Fangyi asked about the third flow (the "Separate" flow). He asked if the "CH-2 IR" block should be shown after the Tx2 as well, meaning the EDA tool is responsible for combining it with the output of Tx2. Walter said no. He said earlier proposals had stated that the model would use the information in the new column (containing "CH-2 IR") but not modify it and not include the information in its output response in column 1. However, now that the new Tx_Impulse_Input Reserved Parameter explicitly specifies whether this new column is required, Walter had intentionally defined the third flow as shown in the diagram. He said the Tx2 AMI_Init would combine the "CH2-IR" into its output in column one. Fangyi said the section defining changes to section 10.2.3, which specifies what the model does with the impulse matrix, did not explain that the Tx2 model is to combine the two input impulse responses in the "Separate" case. Walter proposed adding new language to address this. However, Fangyi then asked if we should swap the two inputs to Tx2 in the third flow. That is, put the "CH-2 IR" in the first column and put the cumulative upstream response in the new column. Fangyi said this would make the flow more like the first flow, except that it contains one new column. Walter said he had created flow three with the intent of mirroring the time domain GetWave flow, and the Tx2 would receive its upstream information as the primary input. He said he preferred it this way, but this was more of a question of personal preference at this point. He also said he preferred to have the Tx2 Init combine both inputs into its final column 1 output, because the Tx2 model would have the best idea of what things looked like at the input to Rx2. Fangyi said that if you just let the EDA tool handle the convolution instead of having the Tx2 model combine the inputs, then you wouldn't need to add new language in the 10.2.3 section. Walter said he would make both of Fangyi's suggested changes for the "Separate" flow, i.e., swap the contents of the two input columns and have the Tx2 model return its modified version of the column 1 input without combining in the new column's information. There were no objections. Walter said the detailed wording for the "Separate" case would have to be changed in the all the flow descriptions, so review of any "Separate" flow language was deferred. Fangyi noted a typo "emmediate" in the Usage Rules of the new parameter. Bob suggested that "immediate" has more of a time connotation, and asked for a different adjective. Fangyi said he had typically used "direct", and Bob said he thought "direct downstream" was clearer than "immediate downstream". Walter said he had rewritten the single-channel reference flows to combine the statistical and time domain flows. He said everything up to and including step 3 is done for both statistical and time domain. Fangyi said that 3.b. was double counting the effects of the channel because it said to convolve the output of the Tx with the channel, but the channel was already the input to the Tx. Walter agreed and noted that section 3 would have to be rewritten. Fangyi noted that in step 4, paragraph 2, the section describing the use of a "unit impulse response", item 2b obligates all models to ignore any crosstalk columns that contain a "unit impulse response". Fangyi said all models would have to be rewritten. Walter said that wasn't the intent. He said the language needed to be clarified, and "models that use the crosstalk columns..." should be changed to "models that use the crosstalk columns to determine the model's equalization..." Walter said he was not aware of any such models already existing. He said they are coming, but they will be new models and can be written to meet the requirement that they ignore unit impulse crosstalk terms. Walter said no legacy models should need to be rewritten. Fangyi said it should be stated very clearly that there is a new requirement being placed on AMI models. We need to make it clear that 7.1 or later models will be required to recognize a unit impulse response passed in as a fictitious crosstalk term. Bob said the version (e.g., 7.1) we are talking about here is the AMI_Version parameter value, not the IBIS version. Bob asked if this would be a requirement for all 7.1 models, or only those that specify the new Tx_Impulse_Input parameter. Walter said only models that use crosstalk information to set up their equalization settings need to worry about a unit impulse response in their crosstalk columns. Walter said if such a model were to advertise itself as a 7.1 model, then it should follow this rule and ignore unit impulse response crosstalk columns when setting up its equalization. Walter added some new language describing how the model should ignore a unit impulse response if it is using crosstalk terms to set up its equalization. Bob said this should go in 10.2.3, and perhaps in two places. Bob suggested that in the definition of the new Tx_Impulse_Input, the descriptions of its allowed values should refer to the flow descriptions in 10.2.3 and also refer to the AMI_Init function signature definitions. This way all the related elements would be tied together in one place. Fangyi again asked if we really need this new requirement. He said the EDA tool could pass in a unit impulse response that was multiplied by a small epsilon. This would keep it from affecting a model's equalization setup decisions. Walter agreed that this was a clever trick, and he said we could add text suggesting EDA tools could use it when dealing with 7.0 and earlier models. However, he asked, "how small is small?" and do we want to make that trick the standard approach for new models too? Fangyi agreed that the new requirement would be more formal and said the epsilon approach was really just tricking the model. Walter said the justification for adding this new requirement on model makers (if their model optimizes based on crosstalk) is that we have a flow deficiency for Init Only models. He said we're usually worried about this issue with Tx models, and if the model maker just made a Dual model then the EDA tool wouldn't be forced to use tricks to avoid deconvolution. So, if a model is Init only and optimizes based on crosstalk, then it should ignore unit impulse response crosstalk terms that the EDA tool might insert to avoid the need for deconvolution. Bob said we should make sure any branches in the flow descriptions capture the fact that if the new parameter does not exist, or its value is "IBIS7.0", then we are using the 7.0 flow (the corrected version - flow 1). Ambrish asked if a 7.1 model without the new parameter would see any change in the regular single- channel flow. Walter said the regular flow is identical. Walter added another item in the 10.2.3 section to address Ambrish's question: 3. Note: when this parameter is not present, or its value is "IBIS7.0", then the normal non-repeater flow is unchanged, except that the impulse matrix may now contain a column with a unit impulse response. Bob noted that column "aggressors + 2" is mentioned but never really defined. Walter said he would draft another version incorporating these changes. Randy and Arpad asked about the timing and whether there was any chance of this being resolved by Friday. Ambrish said we're changing the flow descriptions so we have to review if very carefully. Walter agreed and said there's no chance of meeting the deadline to vote on it at the next Open Forum meeting. - Curtis: Motion to adjourn. - Walter: Second. - Arpad: Thank you all for joining. AR: Walter to create a BIRD211.2_draft2 with today's changes and send it to ATM. ------------- Next meeting: 04 May 2021 12:00pm PT ------------- IBIS Interconnect SPICE Wish List: 1) Simulator directives