This year's recipient of the Kaufman Award is Rob Rutenbar. He has a big connection to Cadence, since he was the founder of Neolinear, which we acquired in 2004. But there is a backstory of how Rob ended up in analog IC design in the first place. And a frontstory, if that is a word, of how there are 25,000 EDA professionals and he trained 50,000 of them. To find out more, I called him up a couple of weeks ago. Early Days Rob, like me, was at university during the "golden era of the rise of VLSI" when Mead & Conway's book came out (more on that later in the week). It became a radical new thing to do a chip as a student, and Rob did a chip of about 2K transistors in 6um NMOS and was immediately hooked. He went on and did CAD-related stuff for his PhD. He joined Carnegie-Mellon University (CMU) in Pittsburg in 1984 since it had one of the strongest groups in EDA, and he started recruiting grad students. In one of those fortuitous strikes of lightning, he went to some dinner event and sat next to Rick Carley. He was a hardcore analog circuit designer (Rob's background at that point was digital and computing). So when Ramesh Harjani asked him what they should work on, Rob said "I like analog, Just met this new faculty member Rick." So they started work on CAD tools for analog. Tools for digital design were exploding, but there was nothing for analog designers beyond SPICE. But, as Rob put it "I was young so didn't know what was impossible." He has ended up spending 25 years building tools. Rick was a great partner since he was working on circuit design. They ended up writing over 50 papers together. He ended up building out the ecosystem—circuit synthesis, rule-based design, large-scale optimization, layout—as part of his PhD, and then with other grad students. In the mid-90s, they were doing lots of interesting design and algorithm work, building out a complete analog design toolset. Neolinear Meanwhile, Rick was doing some consulting on some commercial analog designs. "This is really aggressive," one of his clients said. "How are you going to do it?" Rick said, "We have tools." The tools that Rob and his team had put together. They knew that they should maybe create a company, but they were academics and they didn't know how. But they started to work with some business colleagues who wanted an idea for a startup to use as an exercise. They took the conventional wisdom on how a small EDA company grew (there were a lot in that era, today almost none). The business professors decided to share it with a few industry people without saying who was involved. After all, at this point, it was as much an exercise in putting together a business plan as anything else. So Sony, Toshiba, and others in Japan got to see the business plan: "these (anonymous) academics have these tools and have these capabilities." The response the business professors got was not what they expected. Instead of being an objective assessment of their blinded business plan, the Japanese said: "So Rick and Rob want to start a company? Do it." Rick and Rob founded Neolinear in 1998, with Rob as CEO. That's them in the above picture around that time. After a couple of years they did the CEO/CTO swap and brought Tom Beckley in as CEO, and Ron Rohrer as chairman. So they had CAD-savvy adult supervision. They built out the first version of analog tools. For every step from a sketch on a napkin to the actual polygons on the wafer, they had an offering. Neolinear grew aggressively, took in a couple of funding rounds (including Cadence). It was always obvious to them that Cadence was the right exit, since Virtuoso and ADE were (and are) so dominant. Customers all told Cadence to buy them, and in 2004 they did. The acquisition went well, not one of these ones where everyone quit once they vested, and there is still a big Cadence R&D site in Pittsburg today. Lots of the technology was slipped into the Cadence tools: the core optimizer under composer, ModGen (heavily used for Pcells), the first generation of analog place and route. "Lots of my students ended up in significant positions in Cadence," Rob told me. Of course, the most senior is Tom Beckley, erstwhile CEO of Neolinear, now running the custom IC and PCB group. The CMU and C2S2 Era Rob had taken a year off in 1998 to start Neolinear, and then spent summers there. Gradually it went to a day or two per week, and he was mostly back in academia. From 1984 to 2010, he remained at CMU as professor of electrical and computer engineering. He also founded the Center for Circuit and System Solutions, C2S2, with funding from Intel, IBM, TI, and others. And from DARPA, who were looking at a lot of long-range challenges since everyone knew Moore's Law would end one day. It was a very large organization: 50 faculty, 100 graduates, 15 to 18 participating institutions (not just CMU, but MIT, Berkeley, Stanford, Columbia...). The early work on carbon nanotubes was done there. Illinois One thing that happens when you are an academic, but are also running a $50M research center, is that people call you up to ask you if you'd be interested in being head of their department. One of them was University of Illinois Urbana-Champaign, who invited him to head up their CS department. "But I'm a chip guy" Rob told them them, a bit worried that he had been in IC design for his whole career. "Having a guy heading up CS who could talk to the EE department would be great," they said. So he went to Illinois in 2010. He started to work on accelerator design, especially for speech recognition. "Who does this in hardware?" he asked. "Nobody paints pixels in software, that's what GPUs are for." At that time, speech was all about algorithms for DSPs. So his second career was doing speech-recognition hardware. In 2006, he created a startup (he knew how to do it by then) called Silicon Vox Corp (since renamed Voci Technologies since customers found the "silicon" in the name scary, apparently). A couple of years in, they had serious voice analytics for call centers, including emotional analysis: "on Fridays your agents get sloppy". At Illinois they were working on machine learning. Of course, today everyone is working on machine learning, but they were there early in 2009 since they knew it wouldn't scale in software and needed fast hardware to scale. Pittsburg At the end of last year, the University of Pittsburg (which is not the same as CMU, just in the same city) was looking for a VP of research for the whole institution. He took the position in July 2017, just a few months ago. He covers everything from human subjects, computing platforms, research projects, and more. I guess I should just say "To be continued..." MOOC MOOCs (Massive Open Online Courses) came along in 2011, with Netflix scale video, cloud for automatic grading of exercises, and so on. One of the beacons was Sebastian Thrum's Introduction to AI course. As an experiment, Stanford opened a couple of courses including this one to anyone with an internet connection. To, I think, everyone's surprise, 160,000 people signed up, approximately 10 times the total number of students at Stanford. So in 2012, Rob walked into Stanford, found how the nuts and bolts of how to do it, and proselytized it in Illinois. He found Coursera, and worked with them to put the course VLSI CAD: Logic to Layout online, a course that he had run for a couple of decades at CMU. It was a huge amount of work, but the course was ready in 2013, and ran through 2015. Then Coursera changed their platforms and broke everything (this never happens with EDA software, does it?) so it needed to be revised and by 2016 it was fixed and up and running again. It was now split into two courses, since six-week courses seem to be a better offering for an on-demand course than a whole semester. It still runs on set dates, even though it is completely automated, since it is useful for 500 people to know there are other people at the same point in the same course, struggling with the same exercises, that they can go on forums to discuss. Approximately 60,000 people have been through the course. This is significant, not just for the absolute numbers. When Rob started, VLSI was cool and universities ran lots of CAD courses. But VLSI stopped being cool, and the courses evaporated. Even Berkeley no longer teaches it. It's all machine learning now. But any vibrant field needs an on-ramp. Rob was giving a presentation on MOOCs at Cadence a couple of years ago and was asked what he knew about the people he took it. "I would like to know some outcomes," he said. A Cadence employee in the audience stood up and said that it was responsible for him getting into EDA. If you want to take the course, or find out more, it is available on the Coursera website: part1 part2 . The next courses start on October 30. Hurry, that's today...or another should start in the New Year. Save the Date The Kaufman Award Dinner will be on Thursday, February 8, somewhere in Silicon Valley. Details will be available shortly. Sign up for Sunday Brunch, the weekly Breakfast Bytes email.![]()
↧