AI pcb design

Oct 05, 2026 Last reply: 1 month ago 12 Replies
Jump to replies (12)

formatting link
Maybe AI can do PCB design.



The graph midway down is interesting.


Actually, there are not that many design rules to implement. Should be fairly easy to do.

Cheers

Until you start worrying about cross-talk between traces, and voltage drops along traces.

There were a couple of occasions in my career where cleaning up a printed circuit layout to deal with the unfortunate consequences of these effects turned out to be essential to the success of a project or a product.

The really horrible example was a ground layout which routed about 100mA of logic supply current along a ground trace which subtracted about

100mV from a control voltage that had been carefully derived from 6.2V voltage reference generator. When the control voltage was at low end it's range at it's source it had reversed sign by the time it had got to it's destination.

At least it was easy to fix.

In my experience -- albeit perhaps limited -- the business people really don't care if a project fails because the technical expertise to realize it is insufficient in-house. Nor do they care when people with good technical skills are let go [presumably to a different employer.]

After all, the business people continue to be paid, even when several projects fail.

Best wishes, and happy searching for new employment!

That article is a load of crap. Written by the founder of an AI tool company with a product which assists in PCB design, justifying the need for such a tool.

There is no shortage of PCB layout staff. There is no shortage of engineers full stop. Never have been. Pay enough and you'll get staff.

As for AI in PCB layout the problem is not in routing, but in optimising placement. Autorouters have worked for several decades - if you spend time creating the constraints they work.

As for AI in design it is already offered by the big 3 EDA vendors for both Analogue and Digital work.

Cadence for example have at least four agents that I know of (viraStack ChipStack auraStack innostack) which specialise in different areas. Virastack pairs up with Virtuoso (which is probably used for 99%+ of all analogue chip designs) for analogue work. Aurastack assists in PCB work, and the others in the 1's and 0's stuff.

We've come a long way from when people drew PCB designs with ink on paper, or with black tape on mylar.

I was at a meetup at FLUX.AI last week. They are one of many outfits that are trying to do AI pcb design. They do understand that I want to do some of the layout myself, place and route the critical stuff. But then I want AI to do the grunt work, connect the 150 dangling FPGA pins following my rules and the data sheet, and do the power pours and stuff. Maybe do some thermal analysis.

Allspice.io is interesting. They can use AI to check a design, flag possible gotchas.

Sure, we need to employ PCB layout people, as we need to employ engineers and coders. But we need them to be efficient to compete. We can't just "pay enough" if our competitors can get boards designed cheaper and faster.

John Larkin Highland Tech Glen Canyon Design Center Lunatic Fringe Electronics

Johann 'Myrkraverk' Oskarsson snipped-for-privacy@myrkraverk.invalid skrev: |-------------------------------------------------------------------------| |"In my experience -- albeit perhaps limited -- the business people really| |don't care if a project fails because the technical expertise to realize | |it is insufficient in-house. Nor do they care when people with good | |technical skills are let go [presumably to a different employer.] | | | |After all, the business people continue to be paid, even when several | |projects fail. | | | | | |Best wishes, and happy searching for new employment!" | |-------------------------------------------------------------------------|

Interesting! (S.

formatting link
fuer Kontaktdaten!)

The problem with EDA is always all of the niggly details that need to be present (in a tool or in the mind of the designer) to optimize a design.

Initially, it was petty things like symbol libraries. Footprints. Gate swapping criteria. etc.

Then, things like current carrying capacity of foils based on the static and dynamic needs of the loads they powered.

And, increasingly, there are constraints on trace impedances, trace lengths, etc.

This is all just tedium. Nothing "exciting" to motivate but often consequential if ignored.

Where AI can add value is in extracting those constraints from the components' requirements and the nature of the design and conveying them to a layout agent.

Companies fold if all their projects fail.

Business people do like to encourage their technical experts to make up their minds faster than they ought to. There's a delusion that engineers are pathological perfectionists and won't make up their minds unless put under a lot of pressure to do so.

This usually results in products being released before they should be and the faults corrected by post-production modifications, which tend to be more expensive than a pre-release redesign. On the same project I got to salvage the DC power by replacing a single

2N3055 with a much more expensive single power transistor that could handle the worst case power dissipation. Two 2N3055's would have been a cheaper way to get the power dissipation, but there was no way to fit them into the space available at that stage. Earlier in the process making the extra space would have been trivial.

The power supply design had initially been done by a technician rather than an engineer - a very competent technician, but he hadn't been taught about worst case design. It was a false economy, and didn't do anything good for the technicians peace of mind either.

Well, that part of the design rules. Distance from regulator. Route supplies first, buss's, special, and the rest.

Cheers

We have data sheets that run to 600 pages or more. I'm happy with the ide of an AI tool that looks at my schematic and PCB and prowls all the related data sheets and flags possible errors. Alspice claims to do that and I think it may be useful. They showed mr one example that clains to have discovered an open drain that needs a pullup, and that it checked all the power supplies and in/out direction on all the chips, stuff like that.

Data sheets seem to be getting progressively longer and worse.

That's fine if it's spotting possible mistakes.

John Larkin Highland Tech Glen Canyon Design Center Lunatic Fringe Electronics

formatting link

John Larkin Highland Tech Glen Canyon Design Center Lunatic Fringe Electronics

Exactly. "Not that I don't trust you -- but I don't trust you!" If a technician flagged such a fault, you'd want *him* to back up his claim so no reason why an AI "agent" should be held to the same standards.

It would just be faster, more accurate and more thorough (hopefully).

E.g., the Technical Manual for the MCU I'm currently using is almost

17,000 pages. Wading through that and designing symbolic references for registers, settings within registers, subassembly constraints, etc. is REALLY tedious.

One can trust the manufacturer's opinion as to what the programming model should look like and how it should be conceptualized. But, there's no guarantee that it doesn't contain errors. And, that it fits with your naming and organizational scheme.

If there are N instances of a <widget>, one would think their interfaces would be identical; are they, in fact? How many details do I have to review to convince myself of that fact?

It would be nice to just tell an agent how you want things organized and let it generate the manifest constants, module abstractions, etc. None of this really needs human assistance (other than approving the decisions made by the agent)

Dunno. I've implemented a Production System in my current project and the AI agents just propose new rules -- within the constraints that have been imposed on them (i.e., you can only make recommendations based on observations of the following variables/events). The hope is that this will let a user decide if the "reasoning" behind the rule "makes sense".

E.g., when I was prototyping the HVAC controller, it kept telling me that the air conditioner was broken. At first, I assumed there was some flaw in the code that interfaces to the field -- that some input was being corrupted (after all, the house is *cool*!).

Eventually, I walked through the actual rules with current data and realized where the conclusion originated. Then, manually started making the same measurements.

And, the ACbrrr s*it the bed shortly thereafter!

But, I was able to see where the admonition originated instead of just trusting its conclusion.

Join the Discussion

Have something to add? Share your thoughts — no account required.

Didn't find your answer?

Ask the community — no account required