AI PCB layout

[Cyclotron]

I've been using a ChatGPT-based chatbot to generate PCB layouts in KiCAD.  At least to get the major stuff done, and then I poked at it.
I just saw this article, and it looks like others are much further along than I am on using AI to do PCB layouts.

AI seems to be a popular horse to beat these days.      

[Randy222]

When the work goes into creating the AI models, then that will yield excellent/acceptable results.

That article in-part says:

The folks at Quilter take a very different view. Based on their experience with
professional design teams, they believe this is not how real-world hardware
development works—at least not in the markets they are targeting. In these
environments, companies want their human design engineers to apply their expertise to architect the system, choose trusted components from vetted libraries, and create the schematic.

If that's their sentiment then they will soon be left filing for bankruptcy. The very aspects they mention, can be put into the AI model. No free lunch with AI. The folks with good models will get bigger piece of the pie.

[Berni]

Can you show some examples of the PCB layout that is possible with AI?

Since so far i haven't really seen anything that was properly good. Seen some online platforms that claim they can do electronics design work with AI and i was not all that impressed. It was mostly just feeding netlists trough a general purpose LLM

[Siwastaja]

So... autorouter?

Autoplacement/autorouter algorithms have been a thing for over 30 years, and some of them are actually quite decent when used correctly. It is obvious that utilizing neural networks is exactly the right direction to go into to make already quite decent autorouters even better at it. I'm not surprised at all to hear it works quite well.

Similarly, in IC design (from HDL to masks, or from HDL to FPGA configuration) everything has been auto-place&route since early 1990's if not earlier, manual fine-tuning being a special exception.

[Cyclotron]

Can you show some examples of the PCB layout that is possible with AI?

Since so far i haven't really seen anything that was properly good. Seen some online platforms that claim they can do electronics design work with AI and i was not all that impressed. It was mostly just feeding netlists trough a general purpose LLM

I don't have any designs I want to share publicly, but the article clearly gave one that was spun and produced. Did you look at that one?

[Smokey]

Just thinking about how much of a pain in the ass it is to work with a human "layout guy" that didn't design the schematic but is doing the layout... Even if that dude is a good layout guy it's still painful because they have no way of knowing the design intent of everything.  Even if you did a great schematic and it has lots of layout notes, it's still a lot of back and forth.  This isn't even going into working with the mechanical engineer on connector and mounting placement, housing interferences, cable routing, etc.

So ya, LLMs (or anything else we have now) won't get us anywhere close to full AI Layout (in my opinion).  Autorouters obviously work fine.  But "here's a schematic, let me know when the layout is done" is still fantasy.

[Berni]

I meant more an example of how the whole process goes.

The PCB design process is quite involved with the amount of cross communication required to bring together both the mechanical design and electrical design. So how much hand holding does such an AI need to crank out a good PCB.

If it is AI i would expect it to be a significant step up from the usual autorouters we had for decades. Those are far from "just making the PCB for you" you still need to take care of component placement, you need to feed it a lot of extra information and even then you wouldn't let it just have at it on the whole board (unless you like messy boards) but do it in sensible sections and check that it did what you wanted at each section.

[hans]

AI Clippy: do you want me to fanout that BGA for you?

People ✨somewhere✨ on the Dunning–Kruger curve: Please do!!

https://www.flux.ai/vasy_skral/stm32mp1-linux-card?editor=pcb_2d

AI Clippy: shall I use 10 layers for that?

People ✨somewhere✨ on the Dunning–Kruger curve: YES YES MORE

https://www.flux.ai/vasy_skral/spartan-ai-accelerator?editor=pcb_2d

 

[xplanate]

Honestly, been playing around with ProtoFlow AI lately. You describe what you want in natural language, it pulls real parts from LCSC/DigiKey/Mouser and spits out KiCad/Altium files. way more useful than the netlist to LLM stuff. Still early, but it actually does something concrete instead of just wasting time.

[grx]

and this is why we cant have nice things

is there any craft remaining that will not be infected by "ai" ?

cant wait for the bubble to burst.

[Smokey]

...
is there any craft remaining that will not be infected by "ai" ?
...

Plumbers.  It will be a while before AI can crawl under a sink and replace a p-trap.

[Randy222]

...
is there any craft remaining that will not be infected by "ai" ?
...

Plumbers.  It will be a while before AI can crawl under a sink and replace a p-trap.

Well, maybe the AI issue has escaped you.
If the task is not suited for an AI humanoid bot, then I suspect AI will re-engineer the plumbing configuration so that the task can be carried out by the humanoid.

The other issue,..... once the ML phase has been completed, bots no longer need to learn that task, the learned task is simply shared among the bots. Once one ant knows, they all know.

[Smokey]

...
is there any craft remaining that will not be infected by "ai" ?
...

Plumbers.  It will be a while before AI can crawl under a sink and replace a p-trap.

Well, maybe the AI issue has escaped you.
If the task is not suited for an AI humanoid bot, then I suspect AI will re-engineer the plumbing configuration so that the task can be carried out by the humanoid.

The other issue,..... once the ML phase has been completed, bots no longer need to learn that task, the learned task is simply shared among the bots. Once one ant knows, they all know.

... thats assuming they make plumbing robots.  but in order for the ML to be universal, all the bots would need to be the same, with the same tools.  You can't train self driving on a tesla and expect that to run without modification on a helicopter.

[temperance]

And, at how many dollars is this company being evaluated by some clueless people dreaming about more dollars?

[Randy222]

...
is there any craft remaining that will not be infected by "ai" ?
...

Plumbers.  It will be a while before AI can crawl under a sink and replace a p-trap.

Well, maybe the AI issue has escaped you.
If the task is not suited for an AI humanoid bot, then I suspect AI will re-engineer the plumbing configuration so that the task can be carried out by the humanoid.

The other issue,..... once the ML phase has been completed, bots no longer need to learn that task, the learned task is simply shared among the bots. Once one ant knows, they all know.

... thats assuming they make plumbing robots.  but in order for the ML to be universal, all the bots would need to be the same, with the same tools.  You can't train self driving on a tesla and expect that to run without modification on a helicopter.

If the Hyundai factory Atlas bots know how to build cars, it can also be trained to be the teacher for other non-Atlas bots. Each bot type can identify to other bots it's capabilities, they can all just simply adjust their ways to do the same task.

Bots will be able to share what they know, it's a boom bang process like nuclear fission for each single task shared, except there's no actual kinetic outcome.

You may not see it this June day 2026, but it's coming, and it's coming fast.

[paulca]

You may not see it this June day 2026, but it's coming, and it's coming fast.

At the moment robotics is only really plausible when you have a defined point like a factory or workshop or lab.

So I would have no doubt that "pre-fab" houses could be constructed in "truck sized" components with plumbing already fitted, soldered and what not, by robots.

That will come at lot sooner than a "mobile" humanoid robot that can climb under your existing sink to fix it.

Even in the prefab world, the solution would be to remove that whole kitchen wall and counters and install a new one from the factory.

We don't have bots to do that install yet either.

"Close" in the terms of demos and some industrial and military trials is no where near "close" in terms of the human population permitting them to exist.  If you unleashed a team of bots onto the streets of London I can bet you a months salary they will be found in the Thames in a matter of days.

[tom66]

Reviewed Quilter for work.

- Can't do HDI
- Can't do microvias
- Can't do complex breakouts from BGAs, e.g. diffpairs on alternate layers, without human intervention
- Can't do anything high density (they say generally max 20 pads per sq cm)
- Cost is equivalent to a good PCB engineer ($1/pin), plus you can't trial it on your real data unless you agree to them training on it

So far can't recommend it; but I will look at it again in a year or two and see if they've made enough progress.

[tszaboo]

"tool will automatically select the components, generate the schematic, and perform the PCB layout"
Ah, so it will do like 5% of the job then.
Most of my job is compliance, negotiating with suppliers, DFM, EMC and integration with the mechanical and firmware team. And sitting on meeting explaining that the project is "late" because most of my job is not that 5%.

[tom66]

I saw a good joke the other day.  It said, "as soon as a customer can explain exactly what they want, humans will be replaced by AI".  To which the reply was "ah, we've a job for life then!"

[Cyclotron]

I saw a good joke the other day.  It said, "as soon as a customer can explain exactly what they want, humans will be replaced by AI".  To which the reply was "ah, we've a job for life then!"

reminds me of this meme that went around.