AI PCB Routing Without the Epic Fails: Inside DeepPCB

You guys just mentioned something important, right? The company's been around for quite a while, whereas a lot of the companies we've been talking to lately kind of sprung up over the last few years.

Of course, right as LLMs became popular and now folks are applying LLMs to either code-based or graphical-based approaches to do PCB design.

So how did Deep PCB start? It sounded like it wasn't really necessarily focused on PCB design at first, but you kind of arrived there eventually.

Maybe Alain, Sam, if you could give us some overview on that.

Of course, I can give you a quick history.

So InstaDeep is a general AI company.

Not necessarily an AGI company, because when we started, this word didn't even exist, maybe.

And starting 2017, 2018, we just got ourselves into AI research.

And that AI research was mostly around reinforcement learning.

And that's kind of like the old school deep reinforcement learning where you have all of these agents, you know, DQN, PPO, AlphaZero, et cetera.

It's not the reinforcement learning with human feedback that you're seeing right now in LLMs.

So definitely something that was happening before.

And after a bit of time, of course, we identified the PCB routing problem first with something that looked like it was, of course, very difficult.

It's an NP-hard commutator optimization problem.

And we strongly believed, especially with early experiments and research that we have done, that you could address it with decision making that is sequential.

And if you go into the fancy words, this is kind of like reinforcement learning.

And with time, of course, we published research, we did papers, etc.

But the focus was really during those seven plus years to do the necessary R&D to take DeepPCB to a state where it can be commercialized and most importantly, useful for at least EEs or PCB engineers.

You mentioned something important here, which is, I guess, a sequential approach, right? To me, this sounds like you look at the output, you give it feedback about how it's supposed to change, and then you kind of guide the tool eventually to converge towards a layout that is good enough, right? Meets most of the design constraints, whatever goal you want to have for your design.

Yeah, exactly.

It's a sequential decision-making problem.

Reinforcement learning, as you know, it's trial and error, which is pretty good for, let's say you look at PCB routing, you already have your components on your board.

Then you have a very complex version of connecting the points or connecting the dots.

You have the pins to be connected together.

And in that case, how do you, as a human, how would you address it? You know, you're going to have some maybe high-level strategy of, okay, this has to go around that, and this has to go this way.