On Tuesday AI startup Quilter picked up $10 million in series-A funding to use a combination of machine learning and high-performance computing (HPC) to make designing printed circuit boards a less grueling and manual experience.

While there are automation tools, like auto routers, to assist with PCB layout, Quilter CEO and founder Sergiy Nesterenko argues they can be more trouble than they’re worth.

“They don’t actually understand the manufacturing process, nor the physics,” he told The Register. “They’re just playing a game of connect the dots, and it’s up to you as a user to review their work and determine whether or not that design is reliable.”

Because of this, a lot of PCB design work is still done by hand today – in stark contrast to chip design, which has made extensive use of electronic design automation (EDA) services for decades. Nesterenko explained he became intimately aware of this while developing radiation-hardened avionics for SpaceX’s Falcon 9 and Falcon Heavy’s second stage rockets.

Nesterenko’s goal is to make laying out a PCB more like compiling code. He likened putting together a circuit schematic – a logical representation of how components should be connected to each other – to writing the code, and the actual layout of those components to compiling it. Using its home-grown machine learning platform, Quilter claims it can condense the layout process from weeks down to a matter of hours – reducing costs considerably.

According to Nesterenko, the platform can spit out a complete design while consuming somewhere between $10 to $40 of computing costs. That compares favorably to outsourcing the layout to a third party, which costs anywhere from $1 to $10 per pin – the points where components mount and therefore have to be routed around. Modern boards can have pin counts in the hundreds.

Not another LLM

But, while large language models are generally associated with generative AI chatbots and image generators, that’s not what Quilter has developed.

The model is, instead, based on reinforcement learning – the same technology Google DeepMind used to teach a machine to beat the top human player at the board game Go.

At a high level, the system works by taking a circuit schematic submitted by the end user and feeds it into the reinforcement model, which lays out the PCB design and components while also taking into consideration the physics involved. The resulting layout is then evaluated and assigned a score.

Where things get interesting is how Quilter goes about assigning that score. In order to avoid any undesirable physics interactions – like crosstalk and electromagnetic interference – the model’s work is checked by an HPC physics simulation. This ML/HPC machination is then repeated until it achieves an acceptable usability threshold.

“What we want to do is take that whole process of converting the schematic, a kind of logical description, and outputting a design file or blueprints, that a manufacturer can go and execute on, completely autonomously,” Nesterenko explained.

In other words, the platform should be reliable enough that customers don’t have to second guess whether the design will actually work once manufactured.

The platform is currently in open beta, and users can upload their schematics and test out the system at no cost. While the Los Angeles-based startup claims it will continue to offer a “free forever” tier, it’s also reserving the right to impose restrictions on its use. Once the platform exits beta, Quilter says it’ll “likely charge by use,” and the volume and complexity of the design.

Room for improvement

While Quilter boasts the model can already generate PCB layouts faster than a human being, that’s not to say there isn’t room for improvement. Right now, the model performs best when generating relatively simple PCB designs. We’re told that it’s not quite ready to handle more complex layout tasks, like you might find in a server motherboard.

“The largest board we’ve designed, so far, on the combinatorial side has reached about 2,000 pins,” Nesterenko revealed.

These pins represent components not only have to connect with other components, but also have to route around. Nesterenko claimed Quilter has been able to solve layouts of around 20 percent pin density so far, whereas humans can manage 25–30 percent or better.

Meanwhile, on the physics side of things, he said Quilter has been validated for boards with products like cameras, SD cards, USB, and high speed microprocessors with signal rates up to 500MHz and current up to four amps.

However, Nesterenko emphasized that the platform is making steady gains with regard to these limitations. He added that there’s still a sizable addressable market for customers requiring less complex boards, who can benefit from the faster pace of development. ®

• SEO Powered Content & PR Distribution. Get Amplified Today.
• PlatoData.Network Vertical Generative Ai. Empower Yourself. Access Here.
• PlatoAiStream. Web3 Intelligence. Knowledge Amplified. Access Here.
• PlatoESG. Carbon, CleanTech, Energy, Environment, Solar, Waste Management. Access Here.
• PlatoHealth. Biotech and Clinical Trials Intelligence. Access Here.
• Source: https://go.theregister.com/feed/www.theregister.com/2024/02/15/quilters_ai_platform/