Huang-I apologize for the grumpy tone of my previous reply. The concerns you have are concerns we've heard from other users as well, and the only real answer will be to demonstrate solutions which meet those concerns. Until we have those answers, I admit that I need to be a little more patient.
As regards my statement about "million bits per minute" performance, that was for a full featured time domain simulation using the GetWave function. Runs which invoke only Init() run literally in a heartbeat.
As regards equalization optimization, parameter sweep is definitely not the only possibility. In fact I've used a much more efficient method for years, and I know that I'm not the only one using methods like this. Using any of several approaches, I typically find the optimum combination of transmit and receive equalization for a given link in well under an hour; and combining AMI models with the right combination capabilities in an EDA platform, I'm quite certain that we can find the optimum for complex equalization solutions in minutes.
As regards jitter characterization, while the DJ/RJ distinction is certainly popular, we have had conversations with experts in this area who can't agree on what the definition of RJ is. The main difficulty is that even a process as deterministic as intersymbol interference produces a probability density function whose tails most often closely resemble a gaussian distribution down to cumulative probabilities of 1e-15 or less. That being the case, how is one to distinguish between DJ and RJ? One expert we talked to suggested that the distinction should be made based on the source of the jitter while another expert suggested that RJ should be defined as "that which can be observed to fit a gaussian distribution". My point is that low probability events need to be accounted for even if all that is being analyzed is deterministic processes.
Both of the problems you're concerned about- equalization optimization and jitter analysis can and will be solved in a way that is convenient for practicing engineers to use on a day to day basis. There is not, however, any one simulation/analysis approach or modeling methodology which can be that solution in and of itself. Rather, the solution must combine a number of different approaches and techniques. AMI modeling is one piece of the puzzle, but it is only one piece.
I hope this helps. Mike Steinberger Huang chunxing wrote:
Hi Mike,As a IBIS user, what I am doing it to wonder whether the IBIS model could meet the further requirements of us and I also wish AMI model will be a successful model. I have no intention to offend you or any expert here. In contrast, I really respect everyone's work to promoting development of serial link simulation techniques. AMI model is good solution to many problems which we are facing. But I am still willing to share my opinion about AMI. Althoug, some idea is not so mature. Due to limitation of current encrypted model, we have to omit important components in serial link. Also it is extremely difficult for us to do jitter simulation. We have to analyze the results based on some guess. Encrypted model simulation is time consuming work and the time depends on the model complexity. It's not practical to realize million bit level simulation. Equalization optimization is also another great ability which further help itself succeed. After we discussed it, we only get the parameter sweep method. I don't have any solution to this problem and hope there will be a solution. Parameter sweep is method which depends on the speed of model. What I am afraid is whether AMI model from chip vendors could support this function and the speed of non-LTI model is fast engough. I guess the data one million bits per minute from your model is based on impulse response process method. As for jitter simulation techniques, I still believe that different jitter have different characteristics and requires different method. Sigma level is mostly a factor of RJ and DJ is boundary. It something like TJ peak to peak calculation.TJ_PP@BER <mailto:TJ_PP@BER>=DJ_PP+2*Q(BER)* sI think the sigma level needed in simulation is related with link BER requirement. One may calculate the sigma level from BER through inverse error function. Thanks for your patience and reply. As you suggested, I will shut up my mouth. Regards,HuangHuawei Technologies Co.,Ltd.Tel: 86 755 28976229 FAX: 86 755 28976758 Email: huangchunxing@xxxxxxxxxx <mailto:huangchunxing@xxxxxxxxxx> Web: http://www.huawei.comWarning: The information contained in electronic mail message is intended only for the personal and confidential use of the designated recipient(s) named above. It may be privileged and confidential. If you have received this communication in error, please destroy any and all copies of this message including attachedfiles in your possession. ----- Original Message ----- *From:* Mike Steinberger <mailto:msteinb@xxxxxxxxxx> *To:* Huang chunxing <mailto:huangchunxing@xxxxxxxxxx> *Cc:* C. Kumar <mailto:ckumar@xxxxxxxxxxx> ; IBIS-ATM <mailto:ibis-macro@xxxxxxxxxxxxx> *Sent:* Wednesday, July 25, 2007 9:28 PM *Subject:* Re: [ibis-macro] Re: Query on AMI BIRD Huang- Regarding equalization optimization, it is true that if you have enough time and/or compute resources, you can do just about anything in SPICE, including equalization optimization. The real question, as you suggest, is the amount of time required to get the job done. With our AMI models, we're simulating of the order of one million bits per minute. My experience to date has been that that's fast enough to gain insight by iterating through many solutions. I invite you to consider how long it would take to simulate a million bits in SPICE; and please be assured that, as personnel from both IBM and TI have pointed out, a million bits is required. As regards jitter simulation, part of the question is whether one is concerned with the bulk of the jitter, such as perhaps two to three sigma from the mean, or whether one is concerned with rarer events, such as perhaps seven sigma from the mean (and therefore corresponding to a probability of 1e-12). While time domain simulations may be sufficient for offsets of two to three sigma, they are impractical for offsets of seven sigma. Furthermore, I suggest that attempting to extrapolate from three sigma to seven sigma is not going to yield reliable results. There are other methods, however, that will yield more reliable results for low probability events. I suggest that we've come to a point where we have an adequate mutual understanding, and that further discussion should wait until we have specific results to discuss. In that regard, if you have data you could share with the committee that would help us understand what methods are already yielding satisfactory results for an acceptable level of effort, that would be very useful. Thanks. Mike Steinberger Huang chunxing wrote:Hi Mike and Kumar,Thank you for your both answers.Global optimization is really difficult, or may be impossible. Here is one of the AMI model target, The proposed modeling solution must: - Support automatic optimization of device settings based on simulationI agree that AMI model could support optimization. But, in my opinion, AMI model didn't give much extra advantages on automatic equalization optimization over Spice encrypted model as I expected. If AMI model is fast enough for equalization parameter sweep, this maybe one practical way to do optimization work. Jitter simulation techniques is another problem. Different jitter may have different suitable simulation way. As for DCD jitter, I prefer to the way to modulate DCD on data waveform and simulate it bit-by-bit. Other methods based on approximate algorithmic model may be not so accurate.Regards, Huang Huawei Technologies Co.,Ltd. Tel: 86 755 28976229 FAX: 86 755 28976758 Email: huangchunxing@xxxxxxxxxx Web: http://www.huawei.com Warning: The information contained in electronic mail message is intended only for the personal and confidential use of the designated recipient(s) named above. It may be privileged and confidential. If you have received this communicationin error, please destroy any and all copies of this message including attached files in your possession. ----- Original Message ----- From: "C. Kumar" <ckumar@xxxxxxxxxxx>To: <huangchunxing@xxxxxxxxxx> Cc: "IBIS-ATM" <ibis-macro@xxxxxxxxxxxxx> Sent: Monday, July 23, 2007 4:16 PM Subject: RE: [ibis-macro] Re: Query on AMI BIRD huang: 1. A global optimization probably does not exist. It is even difficult to tell with certainty whether it is better to have equalization at both ends, Tx and Rx. For example in some cases you may find that it is better to have all the equalization done at the Rx end (adaptive dfe , for example), turning on the optimization at the Tx end may even lead topoorer solution.It is best to do a solution space analysis. The present ATM standard should enable you do such solution space evaluation 2. DCD can be applied by the EDA vendor at the TX end and will be reflected in the wave form passed through the GetWave function at theRX.3. Another item for you to notice is the standard per se does nor require you to generate the wave form by the impulse response. in other words the GetWave function is not concerned about where the input wave form comes from. For example if you have a circuit level Tx model , in principle, it is possible to generate the wave form in circuit space and pass the waveform to the RX GetWave. In fact a good(probably at some point a must) sanity test is to verify the algorithm model generatedwaveform against circuit simulation generated waveform. -----Original Message-----From: ibis-macro-bounce@xxxxxxxxxxxxx [mailto:ibis-macro-bounce@xxxxxxxxxxxxx] On Behalf Of Huang chunxing Sent: Friday, July 20, 2007 10:46 PM To: Mike Steinberger Cc: IBIS-ATM Subject: [ibis-macro] Re: Query on AMI BIRD Mike, Thanks for your detailed answer. You really help me know some clues to solve the problems, but I still have some unclear points listed below. 1) What's peaking filter? It is continous time filter composed of zeros and poles? I think the optimal coefficients calculation for non-adaptive equalization is tough task. In my opinion, the impulse response is most suitable for determining equalization tap and coefficients range. When we only consider the transmitter equalization, it may use impulse reponse to calculate optimal coefficients. In order to dsitinguish receiver equalization, I would like to call transmitter equalization emphasis. But when receiver have equalization either, like DFE or CTE, I am not sure the coefficient determined by impulse response will be best. Because the optimal coefficient is only calculated by emphasis itself and doesn't consider the equalization case, like type, coefficients tap and range. I think it will be different from that optimal coefficient calcuated by the method considering emphasis and equalization together. Like two cases below, I think optimal coefficients of case a is different from that of case b. a)emphasis---link---receiver without equalization b)emphasis--link---receiver with equalization My knowledge of equalization is very limited , maybe I omitted some optimal coefficients method which could solve my case, or my talk is just nonsense. : ) 2) I know DCD caused by transmitter is different from that caused by receiver and DCD caused by receiver is most contributed by refernece voltage shift. Here I mainly care about DCD caused by transmitter. Only the DCD caused by transmitter do transfer through channel and will be amplified. Whether the method DCD imposed on the data signal or method which DCD imposed internal to the transmitter model you suggested, I think it should be implemented in non-LTI mode in order to modulate the DCD jitter on data signal. In other words, DCD should be implemented in GetWave() function and simulator will do the bit_by_bit simulation. DidI get the right point?Thanks again. Have a nice day! Regards, Huang Huawei Technologies Co.,Ltd. Tel: 86 755 28976229 FAX: 86 755 28976758 Email: huangchunxing@xxxxxxxxxx Web: http://www.huawei.com Warning: The information contained in electronic mail message is intended only for the personal and confidential use of the designated recipient(s) named above. It may be privileged and confidential. If you have received this communication in error, please destroy any and all copies of this message includingattached files in your possession.----- Original Message ----- From: "Mike Steinberger" <msteinb@xxxxxxxxxx> To: <huangchunxing@xxxxxxxxxx> Cc: "IBIS-ATM" <ibis-macro@xxxxxxxxxxxxx> Sent: Wednesday, July 18, 2007 10:12 PM Subject: Re: [ibis-macro] Query on AMI BIRDHuang- Thank you for your questions. We very much appreciate your interest inthis standard.I have inserted some responses below. Please let me know if I have failed to address your concerns.Mike Steinberger SiSoft Huang chunxing wrote:Hi Experts, I have some difficult in understanding the AMI BIRD. 1) Most equalization of RX and TX is non-adaptive. Chip vendorsdon't have any adaptation codes to embed into DLL function and the DLL function won't have the ability to calculate the optimal equalization coefficients. Then who will be in charge of calculating the optimal equalization coefficients, users or simulators? Is it possible for simulator to do this kind of work?inYou are partially correct in that some of the equalization in the current designs is not adaptive; however, you're not entirely correctthat some of the equalization in the current designs is adaptive. In general, transmit equalization is not adaptive while receive equalization (either decision feedback equalization (DFE) or peaking filter) usually is. Thus, a receiver model truly is responsible for optimizing the receive equalization, and is in fact the only entity inthe simulation which can perform this function, since the receive optimization algorithms tend to be highly proprietary.complex.For (non-adaptive) transmit equalization, the situation is moreIn most real systems, the transmit equalization settings must be determined by experiment during system integration; and it is quitetruethat the transmitters themselves are not able to perform this optimization on their own. In simulation, however, there arealgorithmsgivenwhich can determine the optimum transmitter equalization settingsthe impulse response of the channel and the configuration of the equalizer (e.g., number of taps, tap spacing, available tap weights). The difference between real world and simulation is that in the real world, the impulse response of the channel is not available to the transmitter.AllIn the simulation, a number of optimization algorithms are possible.of them require the impulse response of the channel and the configuration of the equalizer(s); however, the optimization criterioncan vary quite a bit, and so the algorithm to optimize to the chosen criterion can vary as well. Thus, while the optimization algorithm forthe transmitter truly is a separate function, it is a function which offers vendors another opportunity to differentiate their offering. Since the DLL is some packaging that will already be in place, it is also a convenient place to put this other function. Please note that this does not necessarily prevent the EDA platform from also implementing some form of optimization, although to do so, the EDA platform must somehow get a description of the equalizerconfiguration.2)If the channel is LTS, all simulation will be done by impulseresponse mode. How to consider the jitter effects, especially DCD. According to PNA(Phase Noise Amplifier) and experiment, DCD will be amplified by the low-pass channel. Direct statistical processing without DCD amplification won't be accurate.After discussing this point at some length, we concluded that the LTI assumption is not sufficient for all modeling needs. Thus, while the Init() function and its associated processing depends on the LTI assumption, the GetWave() function implements a time domain simulationwhich only assumes that the passive electrical interconnect is LTI. Thus, the GetWave function is free to model nonlinear and time varyingbehaviors of the pin electronics. As regards DCD in particular, there are two entirely satisfactory wayscompliantto model the effects of this impairment using models which arefinitewith this standard:1. The DCD can be imposed on the data signal that is input into the transmitter model. In order to do this, the data signal must haverise and fall times, preferably greater than the DCD to be imposed; however, that is entirely consistent with the real world that is beingthemodeled. Every data generator has a finite rise and fall time.2. The DCD can be imposed internal to the transmitter model. This is probably more appropriate and more accurate since in the real worldDCD will be caused by circuit effects internal to the transmitter. Similarly, a real receiver can introduce DCD as well, and that DCDwillbehave a different effect than DCD at the transmitter. It may thereforeappropriate for receiver models to implement DCD as well.The problem may do not belong to AMI category and maybe I missedsomething here. I hope some experts could help clarify it. Thanks in advance.Regards, Huang Huawei Technologies Co.,Ltd. Tel: 86 755 28976229 FAX: 86 755 28976758 Email: huangchunxing@xxxxxxxxxx Web: http://www.huawei.com Warning: The information contained in electronic mail message isintendedonly for the personal and confidential use of the designatedrecipient(s) namedabove. It may be privileged and confidential. If you havereceived this communicationincluding attachedin error, please destroy any and all copies of this messagefiles in your possession.---------------------------------------------------------------------IBIS Macro website : http://www.eda.org/pub/ibis/macromodel_wip/ IBIS Macro reflector: http://www.freelists.org/list/ibis-macro To unsubscribe send an email: To: ibis-macro-request@xxxxxxxxxxxxx Subject: unsubscribe--------------------------------------------------------------------- IBIS Macro website : http://www.eda.org/pub/ibis/macromodel_wip/ IBIS Macro reflector: http://www.freelists.org/list/ibis-macro To unsubscribe send an email: To: ibis-macro-request@xxxxxxxxxxxxx Subject: unsubscribe