Astrus: Shaping the Future of Analog Chip Design with Artificial Intelligence
Every microchip, even a digital chip, has analog components, as input signals are analog and have to be converted to digital signals first.
While digital chip design surged ahead and is largely automated today, analog design is still a very manual endeavor, taking months to find an optimal design even for simple analog components, which has become a severe bottleneck.
Astrus leverages AI to automate analog chip design, improving analog designers’ productivity by allowing them to focus on high-level, creative aspects rather than performing tedious, manual tasks. Founded by Brad Moon and Zeyi Wang in 2022, Astrus raised an initial round of $275K in February 2023, led by Khosla Ventures alongside 1517 Fund, RiSC Capital, and Raymond Chik. It recently announced a $2.4M pre-seed round led by Khosla Ventures and joined by RiSC Capital, HOF Capital, MVP Ventures, Alumni Ventures, and Plug and Play Tech Center.
Learn more about the future of analog chip design with artificial intelligence from our interview with the co-founder and CEO, Brad Moon:
Why Did You Start Astrus?
Astrus is a result of both my personal journey and maximizing the impact I could have. During my studies in electrical engineering, I specialized in analog chip design, which I saw as a great way to apply physics to something useful for end applications.
For some time, I worked in the semiconductor industry, developing CMOS image sensors for satellites. I experienced first-hand how manual, old-fashioned, and somewhat outdated chip design tools have become. I then spent some time with different startups, but this problem of user-unfriendly, highly manual chip design stuck with me.
Talking to many other analog designers confirmed my original impression: current chip design tools need to be updated. What if I could build new tooling to make analog designers much more productive? Not only was I well-positioned for this, given my personal background, but this could also be a way to massively expand humanity’s computational capabilities.
The amount of computing power globally is a proxy for our ability to solve the hardest problems. If we had 10x more computing power, we would have a lot more capabilities to address, for example, the effects of climate change. Today, we can’t model weather well, but fundamentally, it’s a computational problem – and it’s just one out of many problems that we’ll be able to address in the future if we have much more computing power.
How Does Analog Chip Design Work?
Every microchip, even a ‘digital chip,’ has several analog components, e.g., for the power supply or to receive analog sensor data. Thus, designing a microchip requires typically two separate teams – one for digital and one for analog components.
A digital designer writes code to capture a chip’s digital logic and architecture, which is then translated into circuitry, automatically creating the layout for the transistors on the chip. Digital chip design today is fairly automated; analog design is not.
Take, as an example, an operational amplifier, a common building block of analog circuits to amplify electric signals. Currently, its design involves two key roles: a circuit designer develops an idealized, schematic version of the operational amplifier and runs computer simulations to show that, in principle, it works as intended. In practice, however, a layout engineer has to figure out how to place these, say, 15 transistors, which is a completely manual process and may take up to 20 hours.
The layout is then returned to the circuit designer, who reruns the simulations to find that nothing works – the probability of getting it right the first time is extremely low due to parasitic resistance or capacitance. In reality, wires have length. If you make them too long, you face a lot of resistance, and if you move them too close, you get capacitive interactions between them.
This circle between simulating the analog design and adapting the actual physical layout repeats several times until a satisfying compromise between different constraints is reached – and by then, months have passed.
With Astrus, we fully automated the layout part – so a circuit designer can input their schematic design directly into our layout designer, and we’ll generate physical layouts almost instantly. Finding an optimal layout will still take a few iterations, but maybe only a day – a massive increase in productivity!
You may wonder why analog design fundamentally differs from digital design, so it hasn’t been automated. Digital designs are fairly robust: Even if you have a not-so-optimal physical layout and some resistive or capacitive effects, it takes a lot to introduce errors into a digital circuit, i.e., literally flip a zero for a one. On the contrary, analog designs rely sensitively on the specific absolute value of a signal.
In both digital and analog design, more possible arrangements of transistors exist than atoms in the universe. But while most of those arrangements will work for digital design, for analog design, most of them won’t. It’s a bit like playing a game of Go – there are plenty of options, but very few winning ones.
In the late 1980s, digital synthesizers replaced manual labor in creating layouts for digital design, and there were several attempts till the early 2000s to achieve the same for analog design. But it’s hard – you can’t possibly screen all the possible arrangements to find one that works. With more processing power, it has become possible to train a machine learning model to create layouts for analog designs. It’s now an engineering challenge to make the model good enough, i.e., capture the intuition around which designs will work and what can be manufactured.
Just like AlphaGo eventually defeated humans in Go, picking winning options out of seemingly infinitely many options, we’ll get to the point where AI can do the same for analog design. It will be able to lay out circuits significantly better than any human ever could. From there, we can also move upstream into facilitating the decision-making process and designing schematics. This will change the role of a chip designer from performing tedious, manual labor to focusing on the creative, high-level aspects of chip design.
We are not planning to tape out chips ourselves, as it requires a lot of insight into end applications and their market potential. Instead, we’ll democratize analog design and enable everyone to build any analog circuits, from the parts that go into digital microchips to unlocking analog AI chips and neuromorphic capabilities.
How Did You Evaluate Your Startup Idea?
My thesis around what makes a good startup is to develop a solution to a very specific, high-value problem. Analog design is very specific, as there are only about 60,000 analog designers globally. Yet, it’s very high value and presents a multi-billion dollar revenue opportunity.
We started working with semi-startups, where the founders are both analog designers and decision-makers, so there is no separation of stakeholders, making the sales process more straightforward. When addressing corporates, we’ll win analog designers as our internal champions. With increasing chip demand and a severe shortage of analog designers globally, companies are incentivized to make their analog designers more productive.
What Advice Would You Give Fellow Deep Tech Founders?
Align your personal interest with your ambition. Pick something you’re interested in, and then build a startup in the most ambitious fashion possible. With my background in analog design, I have a huge personal interest in what we’re doing with Astrus – I have experienced the pain of using existing, highly manual design tools. It’s also more ambitious than any of my previous startups, which makes hiring so much easier – people want to work on great and ambitious missions. More ambitious startups are actually easier than less ambitious ones.
Work on the Future of Computing at Astrus
Ready to radically improve global computation? Find out about opportunities to join our Founding Team online* 🚀 🌏 🤖
To reflect this high level of responsibility and impact, we will issue competitive pay and generous stock options.
⚙️ Founding Algorithm Engineer*
As part of the AI team, you will work hand in hand with the CTO (the AI team lead) and product engineering team to ensure seamless integration of AI into the product.
🔬 Founding AI Research Engineer*
Our AI backend is the most important system at Astrus, and the quality of the generated layouts will largely dictate the value of our product to our users and the overall value of Astrus. You will play a leading role in the creation of this AI, the future of Astrus.
*Sponsored links – we greatly appreciate the support from Astrus
Learn more about sponsoring the Future of Computing blog here