Dan Fritchman is an electrical engineer and semiconductor-software entrepreneur. He is the founder and CEO of Generation Alpha Transistor, a San Francisco startup developing AI-based tools for integrated-circuit design.
He earned a Ph.D. in Electrical Engineering and Computer Sciences from UC Berkeley in 2023. His dissertation, An Integrated Circuit Design Framework for Human, Computer, and ML Designers, focused on software infrastructure that allows people, conventional programs, and machine-learning systems to collaborate on chip design.
At Berkeley, Fritchman also worked on open-source electronic-design automation and chip-design education. He helped lead a large student team that designed and taped out a novel integrated circuit within a single semester, which Berkeley described as a first for the university.
His startup grew out of this general research direction: applying modern software and AI techniques to make analog and other complex chip-design processes faster and less dependent on repetitive manual engineering work. Generation Alpha Transistor joined the CITRIS Foundry’s spring 2024 startup cohort.
Tell us about your company?
We’re a seed-stage startup, here to bring AI to the real bottleneck in silicon: analog.
The inspiration for how we do that took a while. I kind of ran my career backwards. I spent the decade from 2009-19 at Apple. Most of that work went into the analog & mixed-signal portions of the iPhone & iPad SoCs. I had the same relationship to that work that I think a lot of designers have: they absolutely love it, and they absolutely hate doing it. I thought probably 1-2% of my time was spent on the core intellectual work. And the rest was really like turning the cranks on a very rickety old machine. And nobody seemed to care much about changing that machine.
I like to point out, if you time-traveled analog designers forward 30 years from 1996 to today, they’d get along just fine. (And maybe the ones from 1986 too.) If you did that to guys in software, they’d think they were on another planet. Manar from Omni Design has a great slide, where he puts the Daisy Systems GUI from 1984 side by side with the version from 40+ years later. And they’re basically the same.
How bad is that bottleneck really? Analog’s become so much smaller of a portion of silicon.
A little illustration of how bad. In my time at Apple we generally made two big chips a year: one for your iPhone and one for your iPad. Call them 2 billion transistors each. So 4 billion transistors a year. When I started we had functionally zero people. By the time I left we had maybe four thousand. So two thousand on average over that decade. So every year, 2k people produce 4B transistors. That’s 2 million per capita. Over everybody. Including the boss, the secretaries, the janitor, everybody. And as an analog designer, my best year was maybe 20k. So two orders of magnitude worse than that average. People say analog is not the majority of the chip, but here’s the thing: if it’s 100x slower, it only has to be 1% of the chip to take as much time as the other 99%!
Back to my backwards career. I wanted to do something like GenAlpha for a long time. I didn’t know how to do the startup thing, and frankly in our field, there really weren’t any. I thought the best place for that energy was in the research lab. I was lucky enough to have great relationships with the best professors at the birthplace of so much of our field, UC Berkeley. So I spent the years from 2020-23 working with Elad Alon, Bora Nikolic, Vladimir Stojanovic, and countless immensely talented researchers, culminating in the PhD. That’s also where I met then Stanford professor Boris Murmann, who’s now an invaluable advisor and advocate for us here at GenAlpha.
So given that history, it’s just absolutely incredible, like my favorite thing in the world, that we have all this new blood and all these new entrants in silicon and EDA. Ten years ago we would have been like the only ones. Now there are dozens. The bad news is, there’s nowhere near as much help coming for what we regard, and I’d say we know, is the real bottleneck – the analog parts. Suddenly the cool thing for silicon startups to say is “we’re taking on Broadcom!”. And they’re right – Broadcom has been absolutely on fire – largely because of their analog, especially SERDES.
What’s your product?
Our first product is an AI Layout Designer. It’s inspired by a very familiar relationship to everyone in analog. It’s sold way more than we had hoped or expected, and in retrospect that makes sense. At every company we hear over and over, custom layout is such a bottleneck. They can’t get projects started because they can’t get them staffed. Even the biggest companies have to ration out the people who know how to use the existing tools, so engineers have to fight for their attention like they’re the prom queen. And much of the work has been outsourced overseas – a fact not lost on our partners in defense and government, and one that makes their lives very hard.
Andrej Karpathy has this great talk and this great metaphor in AI: the autonomy slider. You want products need to be able to ratchet up and down the level of involvement of their human user, or equivalently ratchet down & up their autonomy level. We started with the premise that circuit & layout designers need very low autonomy, and very high involvement. Because they’re used to being in an incredible amount of control. Every wire in a schematic, every shape on a symbol, every polygon in a layout. What we’ve learned since, as we evolve, and AI evolves more generally, is that you don’t need as much involvement. You go from commissioning work that would take minutes, to hours, to days, to weeks. Soon it’s going to be months.
It helps to work in software, the field that’s furthest along in AI, because we get to see the future of every other field. And it’s very normal, right now, for me to commission what would have been months of work, and have it get done overnight. That’s going to be the case for the analog field too, and soon. I tell designers and design managers all the time – imagine how you’re going to do your job when that’s the case. It’s not about speeding something that took 5 minutes down to 5 seconds. It’ll be about writing down what you’re going to do for the next 5 months. Every day.
How are analog designers reacting to AI?
I get to talk to a lot of people in analog about how they see AI. There’s a mistake that many make about where AI’s getting applied in software, and really across the board. They say things like “well we designed state of the art FOM X dBc/ Hz, can AI do that?!?”. And then if you ask “what did you actually spend most of the last 6 months doing”, they looked at their shoelaces and mumble, like, “mostly run the flow and debug the DRCs and do checklists and stuff”.
If you look at where AI’s helping in other fields, it’s generally not like the pinnacle of the field, or the crowning achievement of the best engineer’s career. What works better is, start out by pointing it at the things you just kinda have to do. The stuff your team is never gonna put on their LinkedIn profile. Then you can also notice, this is true in software too. For all the power the LM coding products bring, it’s not like they’re inherently writing better, faster code. They don’t promise “we inherently make Linux 15% faster”. They’re productivity tools. They speed up engineers to try more things. Some of those things end up going more than 15% faster. (There is probably one field which is an exception, where the LMs are actually doing frontier-grade intellectual work, which is math.)
One facet of design work that often comes up in this respect is technology migration. Porting. There are tools for this. But those tools are really rigid. They don’t tend to work, except for the smallest, closest porting problems. You need people instead. Usually it’s like the junior engineer or even an intern. So why is some college student better than a decades-honed optimization algorithm? Because the people have judgment. You expect them to apply engineering judgement. To make decisions like “I should probably pick a different architecture” or “I’m in bad shape here I should probably call the boss”. Those judgments are very tough for traditional software to make. For our AI Designers they’re second nature.
I tell designers all the time: the best way to think of our AI Designers is not as a new kind of tool, but as a new kind of colleague. Like a hyperactive, super-smart intern. For now.
How do customers normally engage with your company?
We will also be at 2026 DAC. We hope to see you there!
Also Read:
CEO Interview with Madhulima Tewari of VerifAIX
CEO Interview with Dr. Il Park of Primemas
CEO Interview with Mohit Gupta of TYLsemi
Share this post via:
TSMC CoWoS versus Intel EMIB Semiconductor Packaging