It strikes me, that AI is rather well placed to do layout of a PCB.
its great at remembering lots of rules,
its great at spacial recognition,
its great at learning by previous examples.

any thoughts,
may be not auto placement for a while, but auto routing seems to be a good AI application,

 

I have not had good luck with auto routing.

 

The AI for PCB design is hype, I've gotten the same email/media spam links from X company. I use standard auto routing all the time for the bulk of PCB design. The key is proper constraints for the things you really care about. This means you need to translate your PCB routing desires into a computer compatible format. It takes time to learn the art for each capable tool.

 

Eagle example:
From this schematic

To this board:

4 layer auto-route. Low effort setting on the computer because this is not the final version.

Top layer.

Bottom layer.

Not as pretty as hand done but it's engineering signal integrity correct.

 

How would "AI" differ then from what the industry already does? they use software already for devising routing. The definition of "AI" has been and will continue to be, variable, changing from one decade to the next.

 

How would "AI" differ then from what the industry already does? they use software already for devising routing. The definition of "AI" has been and will continue to be, variable, changing from one decade to the next.

Exactly. AI means nothing, it's a marketing hype word. Routing is 95% rote slinging connections. The 5% that's left are the EM theory related constraints and controls an 'intelligence' needs to add for proper PCB engineering. There is nothing on the AI front today that can do that 5% better than a human can with the proper experience in design and layout.

 

I don't disagree, that the current routing tools are not very good..
with lots of constrains , on dense boards, and auto routing done intelligently by human driven sets , I have seen good results, but not as good as a human , but more than good enough for the boards they were on .

What I'm wondering is ,
AI , based upon big data models,
would that do a better job.
for instance, when i see a schematic with a DDR4 chip set, I "know" how to route it on the PCB.
when I see PCIe links, I "know" how to route it
When I see DCDC converters, I "know" how to route it.

AI , using the big data models it now is ..
whats stopping AI from knowing as much , if not more,
and then checking my routing,
if it can check, then why can't it do the routing ?

 

I don't disagree, that the current routing tools are not very good..
with lots of constrains , on dense boards, and auto routing done intelligently by human driven sets , I have seen good results, but not as good as a human , but more than good enough for the boards they were on .

What I'm wondering is ,
AI , based upon big data models,
would that do a better job.
for instance, when i see a schematic with a DDR4 chip set, I "know" how to route it on the PCB.
when I see PCIe links, I "know" how to route it
When I see DCDC converters, I "know" how to route it.

AI , using the big data models it now is ..
whats stopping AI from knowing as much , if not more,
and then checking my routing,
if it can check, then why can't it do the routing ?

Where is the domain-specific knowledge data for these big data AI models coming from? This is not social media where there are endless terabyte of English language text/image examples easily accessed 'steal' to make a statistical 'guess' of the next word/pixel. Routing a PCB is not a language problem, it's a process mapping planning problem that big data language AI models don't work very well on. What we need for routing are better 'Expert Systems.
https://www.geeksforgeeks.org/difference-between-ai-and-expert-system/#


https://en.wikipedia.org/wiki/Expert_system

https://bdtechtalks.com/2022/07/25/large-language-models-cant-plan/
Large language models can’t plan, even if they write fancy essays

A study by researchers at Arizona State University, Tempe, shows that when it comes to planning and thinking methodically, LLMs perform very poorly, and suffer from many of the same failures observed in current deep learning systems.

Interestingly, the study finds that, while very large LLMs like GPT-3 and PaLM pass many of the tests that were meant to evaluate the reasoning capabilities and artificial intelligence systems, they do so because these benchmarks are either too simplistic or too flawed and can be “cheated” through statistical tricks, something that deep learning systems are very good at.

The researchers tested their framework on Davinci, the largest version of GPT-3. Their experiments show that GPT-3 has mediocre performance on some types of planning tasks but performs very poorly in areas such as plan reuse, plan generalization, optimal planning, and replanning.

“The initial studies we have seen basically show that LLMs are particularly bad on anything that would be considered planning tasks–including plan generation, optimal plan generation, plan reuse or replanning,” Kambhampati said. “They do better on the planning-related tasks that don’t require chains of reasoning–such as goal shuffling.”

This is really what AI is today.

 

Laying out a PCB manually is easy except for the last two tracks.

The thing that irks me is that the autorouter does the easy part and leaves the last two tracks, which are the most difficult, for you to layout manually.

 

Laying out a PCB manually is easy except for the last two tracks.

The thing that irks me is that the autorouter does the easy part and leaves the last two tracks, which are the most difficult, for you to layout manually.

I learned the cure for that with eagle Topology router. Bull-doze (to a pin, bend or via) a path by deleting tracks and vias blocking the last two. Then restart the autorouter. I have about a good success rate on the reroute.
I also go back and cleanup some the auto-routing that looks too ugly.

 

I learned the cure for that with eagle Topology router. Bull-doze (to a pin, bend or via) a path by deleting tracks and vias blocking the last two. Then restart the autorouter. I have about a good success rate on the reroute.
I also go back and cleanup some the auto-routing that looks too ugly.

Wonder if AI would be able to learn what "looked too ugley"

 

Wonder if AI would be able to learn what "looked too ugley"

I'm sure you could train it on ugly but the real problem is horsepower. The computer resources needed to run current 'AI' methods is immense. You won't be able to make a offline, private DIY design if it required AI processing, it must be pushed to the 'cloud' for processing.
I normally run the Linux version of Eagle on a pretty capable home server workstation. Those little green boxes are the individual cores running for a typical Expert-System auto-routing job. It's not really stressed but a large language routing system would likely kill it.

The guys that run the AI farms will require some sort of payment for making a PCB because they've invest billions to make billions more from jobs like this. We are being JUKED into being dependent on them.