I exhausted the 3 project limit trying to see how it would route open source mechanical keyboard PCBs.<p>Every single one failed with an error.<p>The first one failed with an error that didn't specify the reason, the second one because it had more than 200 wires, though I wonder how that's possible on a 3x3 macropad PCB, and the final one failed with "Sorry, we do not support non 45° rotation for this beta version".<p>The errored projects count towards the limit of 3 projects you can create.<p>Doesn't seem to be working all too well.
Don’t mean to undermine the effort but “ what is the problem being solved here”? I have used Altiums auto router and it’s like Tesla AP version 1. Also like someone mentioned below, the hardest part of PCB routing is Placement. There are Tons of variables that goes into consideration for getting placement correct. Routing follows placement and routing can’t fix placement problems. If cost of routing is what the OP is after then they are mistaken as you could get a 12 layer PCB routed with analog, mixed and high speed DDR and RF within a week by an army of PCB designers in Taiwan and PRC.checkout Palpilot if you need references.
That's human powered for sure. Routing isn't the hardest problem, everything is quick once all the components are placed where they belong.
E.g. TopoR already does quite a good job at the routing part.<p>But if you only take a schematic, and the design is sufficiently complex, it's already difficult if you hand it to another engineer who's not too familiar with what is being done, let alone a machine learning model.
I guess I don't spend enough time laying out tedious-but-straightforwards PCBs to appreciate this. Almost everything I do has some important layout considerations like "this loop inductance should be tiny" or "this section needs guard rings" etc. Also, placing components is non-trivial, if you're doing anything dense, fast, or sensitive. I find that's like 90% of layout work: guess where the components should go, try to route the tricky bits, move the components around a bit, route again, etc...
It looks like a mistake to target KiCad. Pros generally use Altium.<p>2 layers is super limiting!<p>24 hour turnaround is actually quite slow... I would bet they have a human in the loop.
Not terribly positive comments here, but I think people are missing the point. This is an unsolved problem, or at least a problem not solved well. Any efforts in this space should be encouraged. I'm not sure we can trust the main vendors to innovate, and starting off with Kicad makes 100% sense to me at this stage.<p>If this is the first rung of the ladder and it's all up from here, good luck - I'm sure we all hope you nail it.<p>A good comment here about placement - I hope that this approach can grow to adjust placement to some degree, even if not complete placement control.
It is interesting that on the front page, they show a SMT pick and place machine. It only has a small tangible connection to auto routing. It would have been much better if there is a short video showing the routing in action.
Hi everyone, Karim the Co-founder & CEO of InstaDeep here, first thank you for the interest and comments we did not expect we would appear in HN so quickly :) We’ve reviewed carefully the different points made, comments below:<p>“This is an unsolved problem… Yes, this is the point. It's NP-Hard”: "We harness the power of advanced AI and deliver results in less than 24 hours...", which is a red flag” InstaDeep has built credibility in AI circles by innovating in Machine Learning and Reinforcement Learning (RL) specifically on how to find good solutions to NP-Hard problems with AI. For example, you can check our R2 paper (<a href="https://arxiv.org/abs/1807.01672" rel="nofollow">https://arxiv.org/abs/1807.01672</a>) about AI for combinatorial optimization which was accepted at the NeurIPS 2018 in the Deep RL workshop. More recently we’ve just published joint AI-research in RL with Google DeepMind that earned a top 2% global ranking at NeurIPS 2019 (<a href="https://arxiv.org/abs/1905.12941" rel="nofollow">https://arxiv.org/abs/1905.12941</a>). We also have specific domain expertise in Hardware: some of our team members have worked for more than 15 years in this field in companies such as STM, NXP, Dialog, etc. We believe it’s an exciting time to be working on PCB Routing, an NP-Hard problem.<p>“ I would bet they have a human in the loop”: No our system is fully automated (which is why it’s a first) so bear with us as we are in beta :) Automation is possible because we use RL, which is very useful for decision-making problems like PCB Routing. At InstaDeep, we deploy RL systems in the real world and work closely with hardware partners such as Nvidia (we’ve recently been upgraded to preferred partner) and Intel (we’re part of the AI Builder Program). In our opinion, having no-human in the loop is critical to accelerate PCB development cycles, and that’s a key feature of DeepPCB.<p>“Shitty auto routers have existed for 25 years, at least.”: that’s the whole point, autorouters don’t get the job done properly, which is why many boards are still done manually. It makes sense that AI could improve things here.<p>“the hardest part of PCB routing is Placement”: totally agree here, and our goal once DeepPCB routing is out of beta, is to tackle placement. Routing is an important first step, but we don’t plan to stop there. In our livestream (<a href="https://www.youtube.com/watch?v=Ea5i-l8YKQo" rel="nofollow">https://www.youtube.com/watch?v=Ea5i-l8YKQo</a>) a few days ago we clearly mentioned that Placement is our key goal for next year.<p>“I exhausted the 3 project limit trying to see how it would route”: we hear you and have decided to increase the limit to 4 credits per week. We’ve also regenerated the credits for everyone already registered! We will periodically review credits based on demand to accommodate as many users as possible. If you would like more credits please contact us!<p>On a final note, keep in mind we’re in early beta and our goal is to hear from you to iterate and improve the product better. Things might break from time to time and we’ll certainly make mistakes but what matters to us is to keep working hard and make progress on this exciting problem. We believe it’s the combination of Hardware domain expertise and advanced AI know-how that yields strong results. If you agree, don’t hesitate to reach out at support@deeppcb.ai we’d love to hear from you!
It will be interesting to see how this handles signal and power integrity. What design rules is it actually following? If it’s just tossing down routes, AI is not needed for that. Shitty auto routers have existed for 25 years, at least.