[SemiWiki] What are the future technology trend for SPICE Modeling?

CEO Interview: Albert Li of Platform Design Automation, Inc (PDA)

/Original material on SemiWiki/

by Daniel Nenni Published on 11-27-2016 02:00 PM

[SemiWiki] What are the future technology trend for SPICE Modeling?

[PDA] Having sufficient data is really the key to the problem, if data is sufficient, model can be automatically generated or synthesized. The concept has already been applied to the case of passive device modeling, such as modeling inductors. EM solvers play the role of proving more “data” or the synthesizers to generate models automatically. We’ve been working with the same concept for the active devices for quite a while, one way is to enable faster measurements, so that a lot more data can be collected and the other way is to achieve huge amount of data based on limited silicon through machine learning, which requires deep understanding of device behaviors, device modeling knowledge, data for the training and years of training experiences, we have already successfully applied the methodologies to our service projects, and tedious tasks such as model re-targeting is now purely done by machines. Machine Learning enabled model targeting from tweaking model parameters to just defining the targets and let the machine finish the job automatically

[SemiWiki] What are other areas in semiconductor you see that Machine Learning can help?

[PDA] We’ve published 3 papers in the past few years related to machine learning, and we used machine-learning algorithms to help on speeding up soft error simulation of logic circuits, automatic statistical modeling, and automatic RF front-end design,so the areas of machine-learning applications are massive. Algorithms, expertise, data and risk are the four key components to access Machine-Learning applications, take device characterization and modeling as examples, we have been working on the machine learning algorithms for over a decade, and we are definitely the experts in device characterization and modeling, we also have huge amount of data and models from previous projects, and these enabled us to train our software or instrument to achieve faster measurements and automatic model generations.