Semiconductor Insiders - Podcast EP328:A Brief History of Chip Design and AI with Dr. Bernard Murphy

Episode Date: January 23, 2026

Daniel is joined by Dr. Bernard Murphy, a friend and fellow blogger on SemiWiki. Dan explores some key milestones in Bernard’s journey in semiconductors and EDA, beginning with a focus on nuclear physics. Bernard explains how he developed an interest in AI technology and applications. In this broad and informative discussion,… Read More

Hello, my name is Daniel Nenny, founder of Semaywiki, the open forum for semiconductor professionals. Welcome to the Semiconductor Insiders podcast series. My guest today is Dr. Bernard Murphy. Bernard is a friend and fellow blogger on Semiwiki. Welcome to the podcast, Bernard. Hi, Dan, and thanks for inviting you. So, Bernard, tell about what brought you into the semiconductor industry. You know, what's your story?

Sure. So I'll give you an abbreviated version. of how I got here. My academic background is in nuclear physics, a difficult area to make money. So I moved to the US many, many years ago on a CAD group offer from Harris Semiconductor. Outside of undergraduate coursework, I didn't know anything about EDA or semiconductors, but Harris got me started. From there, I progressed through several semiconductor and EDA companies. another startup bug and joined a few, winding up in a Trenta, which I co-founded with Joy Bose, where I conceived of and launched SpyGlass.

I worked in multiple roles, hands-on R&D, R&D management, applications engineering, sales, marketing, and eventually CPO. I was also fortunate to work with several customers who really, really, instilled in me a fascination with distance designs and architectures. I've learned a lot from great people I've worked with in all of those roles. So I think I've developed a fairly rounded view of EDA and how it contributes to design success. After a Trenta was acquired, I considered retiring, but I enjoyed the energy and tech too much to let go. So I talked with a mutual friend, Mike Gianpania, who connected us, and you very graciously agreed to take me on

as a writer to SemmyWiki. So that's the background. Yeah, you've been working with us for quite a few years, and it's been a pleasure, Bernard. When you first started with Semawiki, you wrote mostly about EDA and IP topics, which you still do, but over the last few years, there's been an increasing emphasis on AI topics. Can you tell me what led you to this focus? Sure. So I've always believed that I would be most effective building depth in a few areas rather than a broad sweep. Initially, RTF Linting, because that reflected in my spyglass background,

also adjacent functional verification topics like, stimulation, emulation and prototyping, and then formal verification because I'm kind of a map geek, and I knew this was an area I could dig deeper. Sometimes I would branch out carefully into other select areas that were placed my techie sole. The clock VA, for example, for variance-aware, static timing analysis, and ANSIS for IRDrop analysis. In early 20, In 2020, Paul Cunningham at Cadence, the VPGM of the System Verification Group, asked me to start a blog series on innovations in verification. So each month I would find a topic, and I'm still finding topics, that he and Jim Hogan would add their insights on that piece of research, and I would kind of intro the topic.

That series continues today with now Raoul Campasano joining us after we lost. The innovation series got me looking for interesting papers through research libraries hosted by ICCLE and ACM, and later Google Scholar, because now archive is a major source of research. Since I need to scan a lot of papers to find targets, I developed over time a method to be efficiently selective with guidance from Paul and Raoul. So filtering for intriguing topics and mostly good papers, though I have to admit I've had a few misses, not too many I hope. We cover several AI-related papers for the innovation series, unsurprisingly. And in my search, I also come across papers that aren't really a good fit for the innovation in verification series, but which I find intriguing.

Those have become topics for my own posts, particularly posts on AI unrelated to, obviously, functional verification. AI is very topical, no doubt about that, and also feeds my inner math beak because for a math guy, it's wonderful, it's all linear algebra. These AI blogs are proving to be very popular. I like to think this is because I write these as a non-earned. expert or non-experts. There are a lot of us around. But I also work hard to understand enough to get my brief overviews reasonably accurate.

So that's how I've got into AI. Tell me what you see in active AI applications today in EDA. The question. The production applications I know are mostly under the hood. You're using AI, but you're mostly not aware of it, or only lightly aware of it. Orchestration informal verification was an early starting point.

This manages formal proof generation through a complex mix of using different formal engines, controls, and grid utilization. This previously required manual steering, but through these AI methods, it's now automated. More recently, we're seeing methods to optimize regressions to reduce total verification throughput time and to help with coverage holes in implementation I see optimizing SOC PPA through intelligent

mapping across selected corner cases optimizing system multi-physics in a similar way also planning and optimization for chip-lope-based design now the hot topic there's several capabilities which I see today at a pilot's stage, automatically generating verification assertions from natural language prompts is for me a good example of using large language models to reduce the complexity of using complex EDA tools and flow. So more of that can only be a good thing. The area that gets a lot of press is RTL generation, just like software generation.

So we won't need software developers anymore. we can have the AI do it. I think we are too easily impressed even today that this is even possible. But now there's a big focus because the initial see it can do it exercises were about all it was doing. Now there's a much bigger focus on raising the quality bar for RCL generation along multiple axes. Functional correctness, power, security, readability, and maintainability by an expert human designer.

As judged by these metrics, I see progress with quality still comparable to what software developers report for software generators, which make them promising as a design assistant, but they're still a long way from being hands-free. I believe quite strongly that AI for verification will provide a better ROI than AI for RTO generation. That said, I get the verification doesn't have the media ready sparkle of design creation, which is perhaps why it doesn't get so much coverage. A hot area right now is gigantic flows, especially around regression, triage, and bug isolation.

Again, verification. There are a few very active benches in this area like chip agents, chip stack, which is now part of Cadence and bronco.a.i. This is an area, it might be well worth following. Yeah, I agree with you completely. You know, I think verification is the best application for AI to get started, but I think it's going to go throughout the whole flow. But where do you think we'll see AI advance next in EDA? Good question. So I'm hearing ideas around more usability extensions for complex tools and flows. So the stuff I mentioned earlier orchestrating formal proofs and automating assertion

generation are just kind of the very start of that process. Getting more involved, if you look at many of the tools in the EDA flow, they're tickle-driven. You know, tickle started a long long ago as the way that you could instrument and steer tools. And it's very powerful. It provides access to a wealth of options in these tools. And it's amazingly capable if you know how to use them. But it's daunting if you don't. And there are huge, huge numbers of options.

Instead, imagine saying, forget about how, which particular option buttons I need to push and what values I need to give. Imagine describing a natural language what you want. the AI generate a refined version of that also in natural language that you can polish and you might iterate on that a little bit then it could suggest here is a script to accomplish this task maybe not run it for you because I think there's still a strong feeling that these tools should be assistance not really replacing the engineer and that's because there's the judgment that these go into this process but this is very parallel

with what's happening in software and AI. So it's not exactly the Iron Man's Jarvis, where you just tell you what to do and it goes up and does it, but it could still be a big productivity boost, even for experts and certainly for junior engineers. Incidentally, what I'm describing here in this iterative approach to refining a requirement is building playbooks, which is an idea I find very attractive.

You know, Bernard, just on a side note, do you think AI is going to replace semiconductor designers? Do you think people are going to be losing their jobs? Or do you think AI is going to make people more productive and we can do better chips and more chips, etc? The second one, very definitely. I think, you know, I read an article recently in IA

one of the ICCLE magazines that said even in software, the people who are initially very enthusiastic about using AI to generate software automatically are really backing off of that now and saying it's a good

assistance, it is not a replacement. Just nowhere near good enough to be a replacement. And I think that's a good way to look at it in design as well. Yeah, I agree completely. So Bernard, you've been writing for Semawiki for 10 years since 2015.

You've written a book as well. Now you're writing articles for Forbes. How does that fit in with your work at Semawiki? Very well, I think. It's very synergistic. SemiWiki still provides me with what I consider my primary opportunity, which is to talk to leaders in our industry, learn what they see as trends and challenges,

while also providing me with feedback to test my, particularly my AI views against the technical audience. So I believe very much my opinions have to be grounded in reality with the people who are living this day to day. Forbes extends my reach with what I learn in the industry to a non-technical audience who realized that electronics is a very important part of the economy

but they're much more interested in insights into how what we do or see might directly affect their business if they're in, you know, banking or some other field, retail, for example. I view Forbes as a second channel. So let me with your primary, Forbes secondary, which is a place I can repurpose what I learned through my work with SemiWiki

towards a relatable impact for that more general business audience. I believe looking at, I'm in a group on Forbes, and I believe I have a relatively unique voice in the Forbes Business Council group. I aim to use this to raise awareness in the wider business community around the central role that system innovation plays in AI, but also in other domains, in data centers, in communication, automotive, consumer electronics, IOT, and so on and so on. All the application areas we know and love. So for me, a benefit in working with Forbes is that Forbes forces me to even more to pitch

non-technical audiences. Editors in their review cycle, so I write and they review what I write, don't let me get away with jargon. And they forced me to consider not just to avoid jargon, but why each topic might be appealing outside our highly specialized electronics world. I think that ultimately that's very valuable for all of us to consider that, you know, we're building fantastic technology. but at the end of the day, who cares?

Is this important to, I mean, I know a woman in the street or the business or Wall Street or whoever. So I think it's very valuable. Yeah, I agree completely. Great conversation, Bernard. And thank you again for helping build Semawiki's audience of more than 2 million active readers. You're very welcome, Dan. I thoroughly enjoy working with you and with SemiWiki. And I hope that will continue for a long time.

That concludes our podcast. Thank you all for listening and have a great day.