Learning Curve

4 mins read

Earlier this year, Cadence Design Systems disclosed its ambitions for AI-assisted PCB design. The first phase in its move into machine learning for this sector involves a rollout to a small group of early-access customers.

They expect to use the Allegro X AI software to place components on the layout, route the critical nets and perform signal- and power-integrity checks on the partial layout before human engineers step in to complete the job.

The launch marked another step in the onward march of machine learning in electronic design automation (EDA) in both chip and system design.

In the chip-design arena, vendors have found machine learning to be a useful tool in speeding up some kinds of analysis though there have only been a few forays into tools that can take over tasks that previously would demand human intervention. The most common use-cases involve techniques used to speed up simulations or to determine which formal-verification algorithms would best be deployed on a particular circuit. However, PCB design has seen a small influx of start-ups with ambitions to automate more of the process.

In developing its AI-based offering, DeepPCB took its slogan from the open-source project for chip design initiated by DARPA towards the end of the last decade. Andrew Kahng, professor of electrical engineering at the University of California at San Diego (UCSD) used the slogan “no human in the loop” to describe the eventual aim of the OpenROAD suite of tool on which his group and others worked as part of the DARPA initiative. The rationale driving OpenROAD was to make it possible to take a design in the form of RTL and produce GDS II ready for manufacturing with no direct intervention and do it within a day or so of compute time. Similarly, DeepPCB claims its cloud-hosted system, currently in beta, can deliver complex boards within 24 hours.

The free beta release is for up to 150 components and two layers though DeepPCB plans a commercial version of the tool that will handle larger designs once the beta phase ends, though it will keep the simpler free option to support the hobbyist and maker communities.

For its part, Cadence is focusing on more complex designs, which is one reason why the initial release only performs a partial route. As the export from the AI mode is a regular Allegro file, the designer can use standard rip-up and replace procedures should the automated choices turn out to cause problems later. Saugat Sen, vice president of R&D at Cadence, says the model used by X AI takes account of predicted congestion to avoid backing the human designers into a corner.

A creative partner

Cadence is focusing on more complex designs and believes that is where the demand for AI assistance is coming. “We believe PCB design is getting more complex because ICs are now coming with much larger pinouts. Even for simple to medium-sized PCBs, it can take on the order of two weeks to do manual placement. Doing this in hours is a transformative change,” Sen says.

As found with IC place-and-route aided by machine learning in separate projects performed by Cadence and Google, an AI-based tool can come up with unexpected placement options that are more efficient. “The PCB designer can gain insights from the tool that they didn’t think of before. It can become a creative partner to assist in the design,” Sen says.

Despite an ability to come up with novel layouts, constraints are important, as demonstrated by some users of the online AI tools. Without constraints to prioritise path length, for example, the tools may choose to avoid vias by making routes divert a long way around an obstructing path. That, in turn, leads to changes in how the tools present options to the user as the routing is no longer interactive.

“It is more akin to digital IC design where you have to be upfront in articulating constraints. But somebody who has used Allegro X is already familiar with our approach to constraints,” Sen says. “The longer-term intention is to bring about end-to-end place and route.”

Though a white paper published by Siemens mentions the possibility of end-to-end design, the emphasis at the company as it works on its own machine-learning algorithms to provide interactive assistance.

“Our approach is to leverage AI and machine-learning algorithms to augment the abilities of electronic systems engineers,” says David Wiens, product manager for the Xpedition line of tools at Siemens EDA. “As such, we’re focused on optimising the existing design process rather than ’moon shots’ that in essence auto-generate an entire electronic product.”

Siemens has examined several use-cases for machine learning in PCB design that could speed up design without removing the human engineer from the main tasks of layout and routing. One example lies in the creation of library models, an often tedious job that involves taking the details from data sheets and producing valid symbols and attributes that the layout tool can use to ensure the final design has the correct clock, power and ground routing and fits the proposed layout.

Another possibility that underlines a similar issue with data collection lies in the automation of constraints that will be used to control layout and routing that is at least partially automated. Instead of demanding that engineers re-enter design rules and recommendations for different materials and manufacturing environments by hand, an AI tool could potentially parse the information itself from guidelines and create constraints directives.

Zuken is expected to present some of the fruits of its own work on machine learning for assisted auto-routing in a series of seminars that take place in late spring. The company says it is also working on a form of engineering auto-complete, where the software learns what steps operators often perform in sequence and bring up the most likely commands on menus or hotkeys to reduce the amount of typing needed. Similarly, future tools may analyse previous designs to create design rules for future layouts or determine the optimum distances and routing practices between components in frequently reused types of circuitry.

The use of AI may also make it possible to do more upfront work to determine how changes in materials, production techniques and components might affect the cost and reliability of the final design. Sen says one possibility for the Allegro X tool is to use the speedups it provides to support a larger number of what-if analyses before the team opts for a particular approach.

Celus, another of the AI start-ups moving into this field, tries to avoid its software being seen as an EDA tool. Instead, it was designed to slot into the phase before EDA tools such as Allegro, CR-8000 or Xpedition are normally invoked. The idea behind Celus’ AI-based tool is to take a functional requirement and then come up with a schematic, bill of materials and a basic floorplan that can lead to a working design more quickly.

Although machine learning remains at an early stage in PCB design, it seems set to become a part of the design process. The key question is how far it will wind up being pushed. Will it ultimately take over most of the job or be isolated to problems where it can deliver insights but not take the human out of the loop entirely?