Why are so many startups developing PCB design tools?
Part 2: Quilter, and building a compiler for the physical world
Welcome to part 2 of my series highlighting startups developing tools for PCB design! As we discussed last week, the holy grail for PCBs is designs that are correct by construction. With software, developers are able to describe what they want a program to do in a very abstract language like Python or Javascript, and assume that the machine code that ultimately gets executed performs that desired functionality. In the world of PCB design, there isn’t an equivalent. Designers need to manually go from a high-level design specification, to a schematic with real components, to a physical board layout.
Last week, we talked about how Diode is helping PCB designers get from specification to schematic. This week, we’re moving to the next manual step in PCB design: layout. Quilter, founded by Sergiy Nesterenko, is using AI to automate the process of turning PCB schematics into correct, high-quality board layouts.
PCB layout is a slow, manual process.
The process for laying out a PCB can be extremely difficult, but it’s a very well-scoped problem. As Sergiy told me, the main challenge of PCB layout is simply ensuring that your layout achieves the intent described by your schematic, with sufficiently small imperfections that the board works. “The only thing your circuit board is ever going to do”, he said, “is connect the chips the way you asked and introduce a bunch of imperfections.”
Unfortunately, the physics behind those imperfections can be extremely hard to model. That’s why the most complex PCBs, which can feature high-speed signals or sensitive components, are often laid out by hand by experts. In practice, this means that going from a completed schematic to a manufacturable board can take days to months depending on that board’s complexity. If there are any errors introduced during routing, re-routing the board can take additional time. Given that, it’s probably not surprising that researchers have been trying to automate this process for a long -- over 50 years! But why hasn’t it ever really worked?
Existing auto-routing tools don’t cut it.
From the outside looking in, it’s difficult to understand why autorouting is such a difficult problem -- all you have to do is connect a bunch of wires together, right? Unfortunately, practical PCBs have a lot of physical considerations that existing auto-routers often struggle with.
For example, you may need to design a differential pair, where two wires have the same length. You may need to worry about cross-talk, where electrical signals on one wire can affect another wire placed nearby. Existing autorouters, which focus purely on the circuit topology, can’t understand these key physical aspects of the board design process. Quilter’s technology changes that.
Quilter leverages a physics-aware machine learning algorithm to build a layout engine that can actually understand these sorts of challenges and solve them automatically. Their tool isn’t just thinking about the different electrical connections your board needs to make, but it’s also aware of all of the electromagnetic physics that might cause signal integrity issues with your design. This is important, because these sorts of signal integrity issues are a major cause of board re-spins in practice.
Quilter uses reinforcement learning techniques and electromagnetic physics simulations to train an agent to design boards that take physics into account. It’s more like AlphaGo than it is like ChatGPT; their agent learns by routing designs, simulating them, and seeing where it went wrong. This means that Quilter doesn’t face the same issues that some other AI-for-hardware companies run into with limited open hardware datasets. It also means that their agent isn’t limited to just being as good as a human circuit board designer. They ultimately are aiming to build a superhuman router, which Sergiy thinks could be a major breakthrough for the hardware industry.
AI lets Quilter build the compiler for PCBs.
Currently, the electronics industry is fundamentally rate-limited. The ability to make new hardware can’t go faster than the throughput of a bunch of engineers placing and routing components manually. It’s easy to contrast this with the world of software, where we have powerful tools that let engineers rapidly go from high level representations of what they want to do down to functional software very quickly.
“The thought exercise I like to do”, Sergiy said, “is to think about where we would be in the software industry if the compiler never existed.” Essentially, that’s where we are with hardware right now. Software engineers rarely ever need to write assembly code manually, but electrical engineers currently have to painstakingly lay out all our PCBs manually.
With a superhuman, physics-aware router, all that manual work can be replaced with a tool that ensures that your board works the first time, every time. Instead of carefully manually routing boards to avoid signal integrity issues, and having to re-spin boards whenever you make a mistake, designers can focus on what they do best: designing new, cool hardware.
It’s hard to know what this could mean for the industry, the same way it was hard to see how big of an impact the compiler would have had on the world of software back in the 1950s and 1960s. But it’s exciting to think about. As Sergiy told me, “In software, as a 10 year old, I could make a website or a video game. But soon, as a 10 year old, I might be able to make physical hardware. It’s hard for me to imagine what that’s going to look like, but I find that really exciting.”