Issue link: https://iconnect007.uberflip.com/i/1028393
22 DESIGN007 MAGAZINE I SEPTEMBER 2018 with the introduction of new silicon technologies. Shaughnessy: Outside of EDA, what are some examples of AI that really impress you? White: It is hard to answer that question—there are quite a few. I am impressed with the work Google has done under Jeff Dean. His understanding of the algorithms, and more importantly the implementa- tion details, seems to put him on a different plane than many others. His presentations are mini-courses on problems you should be thinking about if you are not doing so already. Overall, I am fascinated with how far speech processing has progressed with AI/ML technol- ogy fueling it. We played around with neural network-based speech processing for a summer when I was at MIT, and I learned what a hard problem it is …(and I have not attempted to solve it ever since). Google, Amazon and oth - ers have overcome some challenging problems with AI technology, extending it from improved recognition to include translation as well. I am a fan of Netflix and over the years have seen large improvements in their recommen- dation engine. There are some interesting talks that describe the AI and analytics-related inno- vations that go into their ability to personalize and tailor their recommendations, as well as how they tailor the quality of the user experi- ence across many countries and cultures. That is a really dynamic and adaptive environment in which to work and as a user, I see improve- ments regularly. Shaughnessy: Some PCB designers are worried that AI will put them out of a job. I've told them not to worry, but do they have a point? White: Our view is that the introduction of machine and deep learning into electronic design and CAD systems in general will be an evolution not a revolution. In other words, I foresee several stages of innovation as the technology and methodology align. In the first stage, ML/ DL is used inside existing products to improve produc- tivity, performance or quality of results, and we can see that occurring now. In the second stage, machine and deep learning will be used to augment the knowledge of the designer and allow them to work at high levels of abstraction. They can move away from pushing shapes and move toward thinking about system design enablement. It is difficult to capture the intent of the designers, so the challenge is build - ing ML/DL technology that can discover user intent and tailor recommendations accordingly. Think of a design assistant that makes recom- mendations to speed productivity through the manual steps of a design. In the third stage, we use large-scale optimi- zation to drive multiple design decisions to one or more desired states (e.g., in EDA it could be design intent, QoR, or PPA) within a set of constraints (e.g., reliability and design rules). In the fourth stage, we will be able to automate sequences of decisions and eventually achieve longer sequences leading to full flows. Automation will continue to impact electronic design as it has impacted other industries, but I don't see it happening overnight. If we can automate lower level tasks in the next few years, those same designers can spend more time at the systems level focusing on higher- level goals and how to manage mission profiles where single solutions are used or tailored for multiple missions or uses. The lead architects of our OrbitIO solution are working on next- generation solutions to address these goals and have already achieved promising results. Shaughnessy: Thanks for your time, David. We appreciate it. White: Thank you, Andy. DESIGN007 David White