IBIS Macromodel Task Group Meeting date: 08 April 2014 Members (asterisk for those attending): Agilent: * Fangyi Rao * Radek Biernacki Altera: * David Banas ANSYS: * Dan Dvorscak * Curtis Clark Cadence Design Systems: * Ambrish Varma Brad Brim * Kumar Keshavan * Ken Willis Scott Huss Ericsson: Anders Ekholm Intel: Michael Mirmak LSI Amaresh Malipatil Dai Xingdong Maxim Integrated Products: Hassan Rafat Mentor Graphics: * John Angulo * Arpad Muranyi Andrey Matvienko * Vladimir Dmitriev-Zdorov Micron Technology: Randy Wolff Justin Butterfield QLogic Corp. James Zhou Andy Joy SiSoft: * Walter Katz * Todd Westerhoff * Mike LaBonte Teraspeed Consulting Group: Scott McMorrow * Bob Ross The meeting was led by Arpad Muranyi ------------------------------------------------------------------------ Opens: - None -------------------------- Call for patent disclosure: - None ------------- Review of ARs: - Walter send presentation for posting - Sent to reflector - Ambrish send BIRD147 update for posting - Sent draft 9 to reflector - Mike post these - Not done yet, will be done. ------------- New Discussion: Backchannel: - Arpad showed draft 9 of BIRD 147 - Ambrish described the changes. - Training_Mode changes from Value to List Format. - The document is in a more readable format. - Arpad: Are there any IBIS spec portions that need to be declared as changed? - Ambrish: This is all new. - Todd: This is hard to read. - It is missing an architectural overview. - The BCI defines the structure of the passed messages, right? - Ambrish: The BCI gives the minimum the models should support. - The TX and RX may go beyond that. - Todd: This is about standard protocols. - It doesn't say that it emulates the protocol implement in hardware. - It is defined in a way that invites confusion with AMI parameters. - It infringes on AMI In/Out, which tells tools what to expect. - That is now declared here, not in the AMI file. - Does the Training_Done usage rule mean that it appears only in a BCI file? - Ambrish: Yes. - Ken: There was discussion about putting this in AMI, but we decided there could be problems. - Having the TX and RX point to the same BCI file insures they use the same protocol. - We put architectural information in slides some time ago, that could be incorporated. - Todd: The specification needs to be as clear as possible. - Walter: There could be a cookbook for something as significant as this. - Radek: The AMI_Init extension seems primitive. - A new API would be better. - Ambrish: It is only needed when we don't have AMI_GetWave. - Kumar: Radek is asking for a separate API for training. - This is a different architecture. - Vladimir: This is about repeating what is done in time domain in statistical mode. - Why not reverse that? AR: Todd provide BIRD 147 with annotated comments - Walter showed a presentation "Backchannel Issues" - slide 3: - Walter: ASCII files includes .ami and .bci. - These are what the user community would like to see. - slide 4: - Walter: Standards call for using multiple taps, with various minimums and maximums. - This applies to several standards. - slide 5: - Walter: The terms "register" and "index" are important here. - There can be invalid settings combinations. - We don't want register settings, but we do want to use indexes. - Either indexes or coefficients are always exposed. - With statistical coefficients must be used. - slide 6: - Walter: The system integrator must choose presets. - Backchannel uses an associated link which needs BER < 1e-5. - You have to start off with presets that achieve that. - slide 7: - Walter: We need certain AMI Reserved_Parameters. - These are common to all training protocols. - Fangyi: This is for RX to TX adjustment commands? - Walter: Yes. - Fangyi: Would this be in a protocol definition? - Walter: This would be protocol-agnostic. - At the end of TD optimization the TX can be closed and reopened to start normal operation. - We must tell AMI_Init and AMI_GetWave to convert indexes to coefficients. - IC vendors need to supply a utility to convert to register values. - The BCI file really belongs only to the RX. - TX models are very generic. - We might eliminate the BCI file altogether. - I have draft text to put together a BIRD. - It allows either Init or GetWave to handle training. - Would like BIRD 147 to take this approach, or I will propose one. - David: Why does the TX have to tell the tool it is optimizing taps? - Also is there multiple entries into Init? - Walter: It takes only one pass. - Kumar: Real devices are more complex, it won't work. - Walter: I can disprove this. - The EDA tool must know the tap indexes. - Kumar: The TX can send back coefficients at the end of training. - TX buffers can have many architectures. - Walter: We have to solve the system integrators' problem. - RX Init can do everything. - Kumar: An RX can't do that, not in the real world. - Walter: The TX can optimize its taps and can communicate to the RX. - The TX starts with no EQ. - You give it a unit pulse response. - Convolve with the channel and send to RX Init. - Tell the RX what coefficients were used. - The RX can run a simulation inside of itself, applying the TX coefficients. - Init can call its own GetWave to do this, although this may take some time. - Kumar: That's not what real devices do. - The RX decides how to divide up the EQ between itself and the TX. - Todd: Cadence is looking for literal emulation of the hardware process. - That is OK, but there are other, quicker ways. - Walter: With GetWave the adjustments are not made when the RX would actually make them. - Kumar: We need to model the real hardware process and also allow statistical flow. - The BIRD provides that. - Silicon is in the GetWave model, Init is behavioral. - Ambrish: The BIRD has a statistical flow. - It is the most general approach. - Todd: Users need to receive the required settings at the end of the process. - Walter: In some cases the BIOS sets the CTLE. - Some RX buffers auto-adapt and there is no need to report. - The standard requires presets. - Kumar: 802.3 says there must be an initial value. - PCI says presets, IEEE only suggests that. Walter: At DesignCON Huawei showed that training when cold doesn't work when hot, and vice-versa. - One can not rely on training. - Kumar: We have provided the TX coefficients that were requested. - That is a separate BIRD. ------------- Next meeting: 15 April 2014 12:00pm PT ------------- IBIS Interconnect SPICE Wish List: 1) Simulator directives