IBIS Macromodel Task Group Meeting date: 06 October 2009 Members (asterisk for those attending): Adge Hawes, IBM * Ambrish Varma, Cadence Design Systems * Anders Ekholm, Ericsson * Arpad Muranyi, Mentor Graphics Corp. Barry Katz, SiSoft * Bob Ross, Teraspeed Consulting Group Brad Brim, Sigrity Brad Griffin, Cadence Design Systems Chris McGrath, Synopsys Daniel Kirsanov, Ansoft David Banas, Xilinx Deepak Ramaswany, Ansoft Donald Telian, consultant Doug White, Cisco Systems * Eckhard Lenski, Nokia-Siemens Networks * Eckhard Miersch, Sigrity Essaid Bensoudane, ST Microelectronics * Fangyi Rao, Agilent Ganesh Narayanaswamy, ST Micro Gang Kang, Sigrity Hemant Shah, Cadence Design Systems Ian Dodd, consultant Jerry Chuang, Xilinx Joe Abler, IBM * John Angulo, Mentor Graphics John Shields, Mentor Graphics Ken Willis, Cadence Design Systems Kumar Keshavan, Sigrity Lance Wang, Cadence Design Systems Luis Boluna, Cisco Systems * Michael Mirmak, Intel Corp. * Mike LaBonte, Cisco Systems Mike Steinberger, SiSoft Mustansir Fanaswalla, Xilinx Patrick O'Halloran, Tiburon Design Automation Paul Fernando, NCSU Pavani Jella, TI * Radek Biernacki, Agilent (EESof) Randy Wolff, Micron Technology Ray Komow, Cadence Design Systems Richard Mellitz, Intel Richard Ward, Texas Instruments Samuel Mertens, Ansoft Sam Chitwood, Sigrity Sanjeev Gupta, Agilent Shangli Wu, Cadence Design Systems Sid Singh, Extreme Networks Stephen Scearce, Cisco Systems Steve Pytel, Ansoft Syed Huq, Cisco Systems Syed Sadeghi, ST Micro Ted Mido, Synopsys Terry Jernberg, Cadence Design Systems * Todd Westerhoff, SiSoft Vladimir Dmitriev-Zdorov, Mentor Graphics Vikas Gupta, Xilinx Vuk Borich, Agilent * Walter Katz, SiSoft Zhen Mu, Mentor Graphics ------------------------------------------------------------------------ Opens: -------------------------- Call for patent disclosure: - No one declared a patent. ------------- Review of ARs: - Walter email Arpad on status of proposed AMI specification rework - Done - Mike change web page to make items easily linkable - The pages no longer use frames - Next will try to make document URLs shorter - The minutes were broken for several months, but fixed now - Arpad Write a clarification BIRD to discuss accuracy issues related to the various AMI clock_tick algorithms in an IBIS-AMI DLL - TBD - Todd: Update the BIRD for IBIS S-parameter box based on feedback from discussion - No update - Arpad: Write parameter passing syntax proposal (BIRD draft) for *-AMS models in IBIS that is consistent with the parameter passing syntax of the AMI models - TBD - TBD: Propose a parameter passing syntax for the SPICE - [External ...] also? - TBD - Arpad: Review the documentation (annotation) in the macro libraries. - Deferred until a demand arises or we have nothing else to do ------------- New Discussion: Arpad showed the modified AMI Flows document he recently uploaded: - Page 8: - Arpad: Daniel Kirsanov sent comments on this - Walter: Would like to discuss his comments - Daniel makes 3 points: - 1: The flow is too complicated - Left side becomes trivial if every Init returns both combined and filter-only - The same doesn't work on page 9 - Fangyi: How does Tx_use_output fit in? - Walter: Would like to put that off - Fangyi: If Init_returns_filter it returns channel impulse response? - Walter: Or it could return just filter impulse response - Only if it is in the AMI file - I am editing to clarify this - Todd: It will overwrite the impulse response in memory - Fangyi: How do I know how long it is - Walter: That is declared at the input - Fangyi: The Tx equalizer output may be shorter - Walter: Some points will not be used - We should have a control for resizing - Walter: The reference flow will not use that capability - Fangyi: It would be better to have a new reference flow - Fangyi: How do I decide what goes into Tx_GetWave? - Arpad: Tx_GetWave decides on it's own - Todd: Is the dotted line persistent memory? - Arpad: Yes - Walter: The same memory is shared by both - 2: Simulation should not depend on whether GetWave exists - Can't resolve that here - 3: Equalization could be Dirac delta functions - Can't resolve that here - Fangyi: If it is analog the delta T should be given between steps - Walter: The area under a time step has to have a certain value - This is Nyquist theory - Page 10: - Arpad: I will clean this up based on comments - Dotted lines are memory sharing - Solid lines are data flow between functions - Walter: This is really 4 reference flows - Fangyi: 2nd convolution on the right: - Input to Rx_GetWave is Tx_GetWave output without convolving? - Arpad: It's not really a convolution - If we have Tx_GetWave convolution doesn't take place - Fangyi: That is not what we discussed Sep 8 - It should leave the EDA tool to do convolution - Arpad: Yes, if Tx_GetWave exists it does the convolution - Ambrish: That is not what we agreed - Fangyi: It is the opposite of what we concluded - Walter: This flow is complicated - Why not have GetWave do it using overlap and save method - We decided last week that was simpler - Ambrish: Convolution in GetWave might be harder now - Walter: The Wikipedia entry found by "overlap and save" explains it - It is 10 lines of code - Todd: Cadence had proposed Use_Init_output because customers needed it - To date we have not seen a model using this - Any EQ during Init and during GetWave has the entire model functionality - Initializing in both Init and GetWave creates hybrid flow problems - It complicates things but there seems to be no requirement - Use_Init_output should go away and be assumed false - Ambrish: We agreed a few weeks ago and now we are changing - Arpad: It was not an agreement, but a position that could be in flux - Todd: We merely retracted our objection at the time - Walter: We should reset to where we started - Digital input is identical to analog output when Tx is LTI - Ambrish: How will EDA tool handle non-linear parts - Walter: Tx_GetWave could do the linear and non-linear parts differently - Slide 3: - Arpad: Every convolution in the EDA tool has a line through it now - If Tx_Init has no ability to return filter-only it would have to be unconvolved - Unconvolving is a problem - Fangyi: Would be force to do convolution after the non-LTI operation - Ambrish: That is a problem - Walter: Only convolving the filter - Better to have the model convolving so every tool gives the same answer - Ambrish: Why is using this boolean moving the convolution with channel response into GetWave? - Walter: It is not - Ambrish: What changed from 2 weeks ago - Todd: We said the input to Tx_GetWave was digital, which the spec doesn't say - Tx_GetWave did convolution because it was supposedly output of Rx pad - BIRD 107 added Use_Init_Output because Cadence and SiSoft disagreed - Fangyi: Digital was convolved with channel response - That didn't make sense - Todd: We decided to changed - Cadence said vendors wanted Init for peaking filter & GetWave for DFE - Haven't seen this yet - Now we have a digital input to Tx_GetWave - If Tx or Rx can return Dirac filter, why have model convolve GetWave with Init? - Walter: Agree - Todd: Ambrish is asking why not have simulator convolve outside GetWave? - Walter: We still still have to support models without that capability - Ambrish: The question is why are we requiring models to do convolution - Walter: Tx_GetWave applies all equalization to the stimulus - Arpad: Ambrish is asking about convolution with h_AC(t) - Walter: It has nothing to do with that - You always convolve Tx_GetWave output with channel - Arpad: The question is why is it done inside GetWave now - Walter: Error on slide 8 - There should be a convolution in EDA tool in one case - Makes sense if doing clock recovery in Rx_GetWave - Arpad made slide changes - Todd: Model would apply both linear EQ and DFE in GetWave - If Rx has already passed filter why not pass waveform to Rx_GetWave? - Walter: It would be easier for EDA tools with convolution in the model - SiSoft has the code and doesn't care, but the flow would be easier - Ambrish: We should not expect model makers to do more - Walter: We need a list of IC vendors that use Use_Init_Output - Fangyi: If there are no models the flag can go away Walter: Looking forward to AMI enhancement presentation by Vladimir next week Arpad: Are changes made to slide 10 OK? - Walter: Yes - Ambrish: What happened to Rx_Init_Returns Filter? - Walter: It is no longer needed because Tx_GetWave handles the filter AR: Arpad send updated AMI Flows document to Mike for posting Next meeting: 13 Oct 2009 12:00pm PT -------- IBIS Interconnect SPICE Wish List: 1) Simulator directives