March 21-24, 2021
Virtual Conference

Louis K. Scheffer


My working lifetime encompasses almost all of IC physical design. When I was a student at Caltech (BS 74, MS 75), I studied under Carver Mead. This was pre Mead-Conway, and there were almost no tools. IC design was all done with colored pencils and a final hand translation into a new mask artwork language, a recent replacement for rubylith. The only transistors we had were enhancement mode NMOS, with one layer of polysilicon and one of metal.

After graduation I went to work for Hewlett-Packard, and designed a sizeable (for the time) chip, perhaps 10K transistors. The process was a step backwards, with only two layers, diffusion and metal, and non-aligned gates. Everything was drawn by hand, then digitized, then plotted on a giant flatbed plotter. I had to write both a circuit simulator and a DRC program, as these were not commercially available.

Still with Hewlett Packard, I moved to their internal CAD group, where I worked on a variety of programs (simulation, design rule checking, layout, and circuit design) while going to grad school at Stanford. (Thanks, HP!) There I studied under Bob Dutton, who was pushing circuit and process simulation onto mini-computers, and Bill VanCleemput, probably the last guy who could be said to have read every paper in the EDA field. Around this time was my first exposure to commercial CAD, which was design rule checking from NCA, which ran on the corporate mainframe.

By 1981, it was clear that CAD tools could be distributed among engineers and not confined to corporate mainframes. A number of startups, spun out of academia and corporate labs, formed to commercialize CAD. When HP decided to not enter this market, I left and joined a small company named VALID.

At Valid, we wrote and sold schematic capture, simulation, then IC and PCB design, beginning with workstations of our own design. At first this was necessary, but then companies such as Sun and Apollo began to sell hardware that could run CAD tools fairly well. These could be made more cheaply and in higher volume than our workstations, and could be used for other tasks as well. Unfortunately, one of the founders of Valid was from a hardware background and did not believe a software-only company could survive, and therefore insisted on continuing to sell workstations with software. In contrast, other companies began using commercial workstations, which let them concentrate of software, a better use of their time and talent. Eventually, the software only approach won out, and in 1991 Valid was merged into Cadence, one of the many companies to do so.

At Cadence, ICs continued to get bigger, and one by one each issue that had been addressed by hand needed new tools to support larger chips. Floorplanning was needed as the input files to place and route got too hard to specify directly. Crosstalk, voltage drop, and thermal effects needed new tools. Mixed-mode (analog and digital) chips got too big and complex to design with separate simulators, so these needed to be combined. Logic was too time consuming to specify, and had to be compiled from Verilog and then C. All the place and route tools needed to deal with complex manufacturing rules involving more metal layers with different specifications, and every new process contained new restrictions that had to be honored. This is a race that continues today, as this conference shows.

In parallel with these technical developments, the publication landscape was changing as well. Initially, conferences were second class citizens. The “serious” publications, the kind used for tenure and other bragging rights, were the IEEE journals. At first, conference publication could even rule out journal publication, as it was considered prior disclosure. However, over a few decades conferences grew in academic stature. First it was ruled that conference publication would not be considered prior disclosure, and a journal submission was still possible. Then there were special issues of conference results. The ACM put collections of conference proceedings on CDs, and then online, making the articles easy to find. Many specific conferences sprang up - ISPD, IWLS, Tau, SLIP, etc., and became part of the academic ecosystem (published, cited, used to announce new results, became part of a CV, and so on). By 2020, 3 of the 7 most cited papers during 2015-2019 were conference proceedings, racking up several times the citations of the most-cited journal articles. Interestingly, biology is behind in this process, with publication in major journals still the metric of success. In contrast, physics is ahead, with the even faster publication of arXiv considered serious work.

By about 2000, due to the relentless increase in chip size, algorithms developed for IC CAD could often solve problems in other fields, if the connection could be made. Examples I was involved with include SETI@home, which needed band-splitting filters but did not know that the Remez exchange (antique in the CAD world) could design these optimally. Biologists wanted to minimize a linear objective function of placement, but did not know of the (well known to CAD folks) linearization used in the Kraftwerk placer. As biologists began to align huge image stacks, they needed a giant least-squares fit, but were unaware of the sparse matrix routines developed for simulation that were perfect for this problem. Electrophysiologists simulating neurons were unaware of approximation techniques such as AWE, and so on.

Such cross-field communication often led to further opportunities for collaboration. In 1999, I worked with scientists at the SETI institute to write SETI-2020, an attempt to look at the next 20 years of SETI, partially inspired by the semiconductor roadmaps of CAD. In 2008, I took a position at the Howard Hughes Medical Institute, intending to use electrical engineering (EE) techniques to help analyze the circuits of the brain. But I soon found that there were almost no detailed circuits to analyze, so I started working on brain reconstruction techniques. These have improved to the point where we now have the circuitry of the central portion of the fly brain, and I’m back to studying the operation. During this time, I’ve tried to bridge the EE-biology gap by presenting biology perspective at EE conferences.

One lesson I’ve learned, and would like to share, is that cross field collaboration is easier and more fruitful than you might think, for the state of the art advances unevenly in different fields. What you think is routine may be a big advance in another field, or conversely someone in another field may have already solved a problem analogous to one that is vexing you. And socially, in my experience, folks in other fields are never offended that you are poking around in their area of expertise, instead being happy someone new is interested, and welcoming any help they can get. So if you’ve got an interest in another field, read some articles, go to a conference, and ask some questions. You’re bound to find it interesting, and your experience may be more valuable than you might think!