Issue link: https://iconnect007.uberflip.com/i/666328
68 The PCB Design Magazine • April 2016 AMI models without having to modify the de- sign data in any substantial way. Too often what happens with IBIS-AMI mod- els is that the model development only takes place after the SerDes design is done and the glue is dry. The IBIS-AMI model developer has to create a model that describes the SerDes be- havior by literally redeveloping it from scratch. It's not a development process; it's a rede- velopment process in the classical sense. The question has always been, "Can we take a snap- shot of the current SerDes design at any point in the design process, and without rewriting it, use it to generate C code and make AMI mod- els?" What we're presenting with MathWorks is exactly that, using the combination of Simu- link and MATLAB code that are characteristic of SerDes design to generate IBIS-AMI model code. The generated code is compliant with the IBIS- AMI interface and runs in commercial IBIS-AMI simulators. There are a number of interesting things that now begin to happen. You can say at some intermediate point in the SerDes design cycle, "Let's give this model to some key customers to see if it's going to work for them." This gives you the ability to take critical SerDes design de- cisions and find out if they are going to work on the types of channels your customers are de- signing in their next generation systems. Nor- mally, the only way you know that is by build- ing silicon and giving it to a customer, but by then it's too late to do anything about it. If we reduce the effort needed to create an AMI model down to the point where SerDes de- signers can create AMI models in real time dur- ing the active SerDes design cycle, it becomes possible to give intermediate models to custom- ers, who can say whether that design will work for them. If it won't work, those same custom- ers can use that intermediate model to identify what changes are needed—more taps, different adaptation loop bandwidth, or whatever it hap- pens to be. That, we think, is a game-changer. One other key point in developing good AMI models (an AMI model always has two parts): an analog part that models insertion and return loss, and an algorithmic part that mod- els equalization and clock recovery. You need to design the two parts of the AMI model so they work together to produce the correct overall re- sult. We see too many AMI models where every- thing has been put in the algorithmic model, and the analog component is "idealized" with an ideal resistance and no capacitance. That's just wrong—there's no way a simulator can cor- rectly predict reflections and ringing if the ana- log model doesn't correctly reflect the silicon impedance. It comes down to this: If you don't get the analog model right, you don't get the insertion loss and return loss right. If you don't get those right, you don't get ISI right. If you don't get ISI right, you don't get the waveforms right, and you don't get the eye diagram right. That's the knock-on effect of not handling the analog model properly. The flow that we put together with Math- Works basically ensures that your analog model is correct as you start building the algorithmic behavior, and allows you to refine both of them over time. They're always being evaluated to- gether, so the combined model is going to work correctly. You're going to have the right analog behavior, you're going to have the right algo- rithmic behavior, and the two of them together are going to give you the right model. We think it's a big deal just because we know how many AMI models we've gotten from other people where the analog part is compromised and has affected the simulation results. Shaughnessy: Well, that sounds really promising. Todd, thanks so much for sharing this with us today. Westerhoff: Thank you, Andy. PCBDESIGN sisoft: optimizing the state of the art FuRTHER REaDIng 1. To attend the webinar "A SerDes Balancing Act: Co-Optimizing TX and RX equalization Settings to Maximize Margin," click here. 2. For more information on Optimeye, click here. 3. For an Optimeye evaluation request, click here. 4. To downloand the technical paper "new SI Techniques for large System Performance Tuning," click here.