DAC 2019: Day 3

Two keynotes get day three of DAC started. The first by John Cohn, Massachusetts Institute of Technology & IBM Watson AI Lab. “I am a nerd. Look back 100 years in processing. We have gone from mechanical computing to where we are today, but it has not been a smooth curve. There are smooth places and then discontinuities. This is when what you were working on no longer works. How we make those leaps is what defines us.”

He then migrated to personal disruptions in our lives. “How do you face those? Do you sit back and wait? Sometimes you get to make those decisions. Other times they are made for you.  What makes us resilient has a lot to do with ‘play’. When you play you are in the zone and you are experimenting and open to new ideas.”

Play drives innovation and it is important to keep an aspect of this in your work. But you can’t predict the future. He told a sad tale about the death of one of his sons and the way he dealt with the grief was to start to play. “I love making things and worked with my hands to heal. It helps you find your tribe. I started to take technology and sharing it with people. When you share your passion, it comes back to you. A tragedy changed my focus and I started working on IoT and that was when I fell in love with Artificial Intelligence.”

He then discussed issues such as privacy and the negative aspects of the technology. He talked about making technology more approachable.

“It is time to misbehave. Nobody got to where they are by doing what they were told. Play is a great inoculation for worry. We grow old because we stop playing.”

The second keynote was given by Bas Verkaik of SPIKE Technologies. He and a group of other students created an electric bike and rode it around the world. He started with a video about their journey, the problems, the excitement and the reception they received.

“We were a bunch of students and had not even finished our bachelor programs. An electric bike capable of going around the world in 80 days did not exist, so we had to develop it. We were ambitious but had little experience. The dreams within the team needed to get sorted out.”

He described their design process and how it evolved. They had to build trust as they progressed. They had failures and had to reach out to others for help. They finished up getting help from 150 companies. “People all over the world became part of our team. We had huge crowds waiting for us. It inspired people.”

After a short coffee break, Chris Rowen, CEO of Babblelabs, presented ‘Deep Learning Meets Silicon – a progress report on technology, applications, startups and challenges.’ “The core idea of deep learning is that we are going to build a complex mathematical model that mimics the behavior of an extremely complex but hidden system. We do that by getting a lot of examples and labeling them. We run them through the model and see what we get. We nudge the coefficients in that model to get the right results. It is a very computationally complex task. But this model can be applied to almost all types of data.”

He talked about the statistical approximation techniques used to solve many types of problems. “It is having an important impact across an increasing number of applications.” Then he started to get into the mathematics. It is called deep learning because the number of layers of linear computations can get to be very deep. It defines the number of computations that you will need to perform and potentially the rate at which it has to be performed.

“Memory is dictated by the size of the model and the number of coefficients. But the code size is often quite small. This highlights an important engineering issue – we tend to fixate on the evaluation itself. It may represent 99% of the computational horsepower, but all of the other things around can be much larger in terms of code size. The software development is only partly limited by the neural network engines, but by all of the other tasks around it.” More coverage on his talk in the future.

I had a working lunch today as I moderated a panel for Cadence on the subject of optimizing verification throughput. Panelists were from Arm, Intel, Samsung and Cadence. The great news was that they were all willing to share their experiences and the paths that they were taking to improve their verification effectiveness. We will most certainly be bringing you the highlights from this panel.

Following lunch, I sat in on another panel titled “Should we trust AI for Cybersecurity?” Would they persuade me that my initial answer of – you must be crazy – is naïve? Panelists Brendon Dolan-Gavitt – New York Univ, basically said that we need a stronger grounding before we trust it. Lejla Batina, Radboud Univ said we must find a way for AI to become trustworthy, because it is here to stay. Rosario Cammarota, Intel concludes that any example that uses real time data is not yet ready for AI, but there is hope. And Bita Rouhani, Microsoft Research thinks the lack of an “I don’t know” label is problematic.  Dolan-Gavitt summed things up nicely saying that AI is better as an attacker than as a defender.

The last panel of the day posited the question – Secure Open-Source Hardware: Hype or Reality. The premise of the panel is that we have seen a reduction in trust of our CPUs ever since side-channel attacks started to be uncovered. By opening up the CPU architecture, does it become possible to do a better job? The introduction also focused on the languages and tools required to support the open source community. This is a subject that we will be writing about in the future.

While this concludes my write-ups of DAC, Semiconductor Engineering will also be at the Thursday keynote and will bring you more complete coverage of all the keynotes.

Check out what happened Monday in DAC 2019: Day 1 and Tuesday in DAC 2019: Day 2.