DeepPCB beta: TensorFlow-based machine learning PCB routing, online free trial

[I wanted a rude username]

The results it produces are ... interesting ... and it can't do placement yet, only routing ... but it does work at least some of the time. https://www.deeppcb.ai/

Launch video (not much content).

Hacker News discussion. Best comment: "Yeah, but that failure was the result of advanced pure AI cloud native deep learning neural networks. The future is now. In a few more years you won't even be able to tell if a person gave up and quit a trivial task or if it was just a sophisticated machine."

Seems like this isn't particularly useful at present, but if they add placement (and of course permit more complex boards) it could be a game-changer. Thoughts?

[mikeselectricstuff]

Like normal autorouters, it will still take a lot of time to set up all the constraints of your particular design. More so once you add placement.

And how do you tell it about nets that can connect to any one of x pins etc. ? Does any existing system do this even?

[SiliconWizard]

No quantum computing though?

[texaspyro]

No quantum computing though?

And no blockchain?      It ain't swoopty noodles unless it has blockchain.

[chickenHeadKnob]

Seems like this isn't particularly useful at present, but if they add placement (and of course permit more complex boards) it could be a game-changer. Thoughts?

Placement and the signal conditioning that go with it are  hard problems. Give a mixed signal schematic to a technician level layout person and you are likely to have analogue signals routed right through a digital thicket on the way to the A/D on the far side of the board. This scenario actually happened on a project I worked on once, where my responsibility was only for the software. The engineer and PCB layout guy could have asked me to review it before sending it but no the layout guy was really proud  of his work and  boss's favourite brown-noser. Expensive boards were trash.

[SiliconWizard]

Jokes aside, a tool that would really leverage AI to do automated placement would be more useful IMO. Any decent modern autorouter with the right set of rules will do a decent job at autorouting if the placement is adequate, but good placement is key and pretty hard, and time-consuming to get right.

Regarding DeepPCB itself, this just looks like the very typical small start-up. Their current model is to "eat" the most PCB designs they can (as far as I've seen taking a quick look, the only means of trying it at the moment is to submit your PCB to them through their web site and wait?)

[thinkfat]

Has anyone actually gotten any result out of their router? I've give it two of my simpler, already finished projects just with all traces ripped up, with a known-working placement. Each of them failed to be processed even to an intermediate result.

I call this a fake.