While anyone can learn how to design a circuit board, it takes a skilled engineer to design a circuit board that is both well optimized and unlikely to melt, explode, or cause whatever it's controlling to melt or explode. Skilled engineers tend to be busy and expensive and on the ornery side, especially if you ask them to do things that don't take full advantage of how brilliant they are.
JITX is a startup founded by a team of electrical and mechanical engineers from UC Berkeley that's building AI with the goal of designing optimized circuit boards in hours instead of weeks. As a result, the engineer’s hands-on role is minimized in favor of a more supervisory role. You tell the system at a high level what you care about, and then it designs a working PCB with everything you don't care about figured out for you. Your engineering expertise remains focused exactly where it’s needed, and the system turns out circuit board designs that are faster, better, and cheaper.
The CEO of JITX is Duncan Haldane, who is responsible for a “hyper-aggressive pogo-stick” robot called Salto-1P. A substantial amount of the work that went into designing Salto-1P (as well as many other robots) involved developing custom hardware, including circuit boards. Haldane and his colleagues found themselves having to continually start over from scratch every time they wanted to do something new. “We realized how little of our time we were spending on our core activity—the research,” says Haldane. “All of our time was focused on low-level hardware design. The amount of effort it takes to design custom hardware is a huge barrier in the way of new and creative systems.”
JITX's goal is to make hardware development more like software development. It’s not the first company to try something like this, and there are already plenty of tools out there that can assist with circuit board design in various ways. Nonetheless, JITX feels that its more comprehensive and holistic AI-driven approach is unique. From the company’s announcement today:
Our core technology is inspired by the technology used for designing computer chips. The introduction of Hardware Description Languages in the 80s revolutionized chip design. HDLs fundamentally changed how engineers designed circuits. Instead of manually drawing the shapes that make up the circuit, engineers would instead express the intended behaviour of their circuit using code, and then have algorithms automatically translate that code into the necessary copper shapes. This workflow is what makes possible the billion-transistor chips we see today. We bring the same workflow to PCB design.
Circuit board design is a multidisciplinary challenge, and we have to factor in electrical engineering (circuit design, RF design, signal and power integrity), mechanical engineering (thermal, vibration), and manufacturing (cost optimization, DFM/DFA/DFT). Unsurprisingly, almost every subproblem is computationally intractable, so we use clever representations and learned heuristics to arrive at good solutions. There are a million details to keep track of across all of those disciplines, and it’s high time we get computers to do the bookkeeping.
Here's an early demo of JITX that Haldane recorded back in November of last year, to give you a sense of how everything works together:
“What they're trying to do is something that most hardware people want to see,” says Ted Larson. Larson runs OLogic, an embedded systems research and development company with a focus on robotic applications. OLogic has worked behind the scenes on many robots that you're definitely familiar with, none of which they can tell you about. “One of the things that they've identified, that I totally agree with, is that Silicon Valley has gotten notoriously horrible at building hardware,” Larson says. “The premise of, doing hardware is hard, and there’s a huge shortage of people doing it inexpensively? That is spot on.”
There are reasons, though, why the hardware-as-software approach has yet to pan out, Larson says. “There's a desire in the world to take electronics and make it more like software engineering, but to do that with circuits, beyond what you'd just use for prototypes, becomes challenging.” Designing for compliance testing and manufacturability requires experience, and so does selecting all the right parts that will work together in exactly the way that you want. Larson suggests that JITX might be ideal for projects that are somewhere in between the prototype stage and making production electronics, especially considering the savings in time and cost over more traditional approaches. And there's certainly potential for more, says Larson. “They're really early in this process, and there are enormous opportunities for them to make this process better.”
At the moment, JITX uses the tools that it has developed internally. You tell them what you need a circuit board to do, and they'll get human engineers, augmented by their AI, to tackle the problem and get you the board you want more efficiently. On average, JITX produces circuit boards three times as fast and 25 percent cheaper than experienced humans working unassisted. The end goal is to increase the automation even more, and expand it beyond just circuit boards. But for now, JITX will be joining Y Combinator's summer class to develop its ideas with a little extra support.
For more on JITX, we spoke with CEO Duncan Haldane via email.
IEEE Spectrum: Can you describe the conventional process for designing a circuit board? What sucks about it?
Duncan Haldane: The first thing you do is figure out what the circuit board needs to do: What are my actuators? what are my sensors? how is it going to communicate? how much processing do I need? From there on in, it’s a grind.
You Google around and find some perceived optimal set of parts, digging into the 100-plus-page data sheets to see how to make everything work together. After that, you model all the parts in your favorite CAD tool (by manually transcribing information from a PDF), and then draw a schematic. In the schematic, you add symbols representing all the parts on the circuit board, try to arrange them nicely (in 1 page or 50 pages, depending on the complexity of the design), and manually connect pins together by drawing little lines that represent wires. Hopefully you don't make any mistakes when drawing the lines. At this point you call in some other engineers to stare at the schematic for a few hours to try to find the mistakes, because a single mistake means your board doesn’t work (and might explode).
Once you have the schematic, you can then design the circuit board itself. The goal is to figure out where to put all the components so that you can draw in copper shapes that will connect up all the pins. There’s a whole load of physics to worry about, and you deal with it by running some simulations and manually mapping the results over to your board (by drawing copper that is the right shape). At this point, you also need to worry about how the board is going to be manufactured, assembled, and tested. You call the engineers back in to stare at the board design for a few hours to try to find more mistakes. Hopefully they find all the mistakes, because now you're ready to pay for your circuit board to be manufactured.
What sucks about it? Everything. It's the worst human endeavor. Imagine going through all that and then starting from scratch on the next design.
Why hasn't someone already done what JITX is doing now?
This is happening now because the industry need has become desperate. Circuit board design is now so intricate that human teams hand-draw graphics in shifts 24/7 to meet deadlines. Furthermore, new factories are highly automated, making design the bottleneck between companies and the market. And this urgency comes at a time when AI is wiping the floor with previous state-of-the art methods. We’re the ones connecting the new results from AI to this deepening problem in the market.
How much of the circuit design process can be automated, and what would that look like to an end user?
All of it can be automated; you just have to have the right approach. What it looks like to an end user is very much the way our tools work today. The user says what they care about, not how. For example, we request a board with BLE and a microphone (the “what”), and our software selects matching key components from the library, solves for power supplies and component values, sources all the parts, assigns pins, plans out placements, routes traces, and then exports the board and schematic (the “how”). If you care about the shape of the board, add it as a constraint, if you care about the position of a component, add it as a constraint, if you know which BLE chip you want, add it as a constraint. Design tools should be smart enough to solve for the million details you don’t care about, and optimize your design for what you care about.
What kinds of constraints will there be on the system you’re working toward?
The biggest constraint is that designers give up control over the small stuff in exchange for automation, correctness, and speed. It's a big change from the way tools work today.
What are you working on with DARPA?
We’re part of the Electronics Resurgence Initiative (ERI), a massive collaborative effort that DARPA is funding to reinvent how electronics are made. Specifically we’re in the IDEA program, which aims at no-human-in-the-loop design of electronics. For IDEA, we’re working with semiconductor and component companies to build out a vast library of electronic components, creating software that turns high-level design goals into real circuit designs, and creating optimization tools that find the best possible version of the board you want to build.
What's your longer-term vision, and how does what you’re doing now factor in?
Longer term, we want to scale out this automatic circuit board design technology so we can free up hardware designers to be more creative. Everyone gets on-demand custom circuit boards, and we make sure the problem stays solved. Afterwards we want to get back into full electromechanical design. A lot of the difficulty in designing a circuit board is negotiating with the mechanical engineer on how to get an extra half millimeter of room so you can fit that one component in. Robots are electromechanical devices and we’re working towards the day when our software can design better robots, which will assemble the next generation of even better robots, running code on even better circuit boards, and so on.
Today, we’re dogfooding our design tools—using them ourselves to quickly design awesome circuit boards for other companies. It’s the best way to make sure that the tools work, are usable, and that every minute of development is well spent as we keep our eye on the prize.