We have 240k users on the platform…hobbyists, students, small/mid sized teams to Fortune50 companies. We host millions of projects!
It’s a huge ecosystem not just of users but also semiconductor manufacturers, fab houses, etc.
Think GitHub on steroids!
And it’s the only AI first end-to-end design tool in the market…you can go from idea to research to architecture to BOM to schematic to PCB layout and then to manufacturing. Workflows that usually take months in other tools can be done in a few hours to minutes in flux.
It’s still early days though but we are moving insanely fast! So if we don’t met your teams bar today I suggest checking back in next month.
I think it’s good for more players to enter the market…EDA tooling is stuck somewhere in the 80s and we need to fix that.
It’s too expensive and risky to make hardware…and we deserve better!
Silicon is the most important resource in the world, as evidenced by FAANG
FAANG’s ability to scale up compute is limited by their ability to design, test, and manufacture silicon. Every competitive AI company is working to build it’s own
Everyone is focused on making software layer more efficient, we are seeing 3-5x productivity gains
I am the founder of https://flux.ai and we are the only ones making the hardware layer more efficient, 6-15x productivity gains!
As you saw with NVIDIA’s growth, the hardware side of this is as valuable as the software side
Flux.ai is a AI first hardware design platform build for the modern age. Our mission is to take the “hard” out of hardware
It's why we never implemented it in KiCad, and why Blender has a "2d drawn navigation controller" instead of 3d drawn.
But yea, I'm not the biggest fan of the Flux approach but maybe I'm too traditionist. I'm a big fan of _understanding_ the design. Yes, I can even use a code gen to generate a fuck ton of boilerplate now with AI. But do I actually know what it's doing? At a surface level sure, at the indepth level? Ugh I don't look.
The penalty of using AI at hardware is the designer loses any forced motivation to actually understand what they are designing. This is critical to be able to troubleshoot designs, and even do things like worse case failure analysis and anything fancier.
For simple designs, sure, you probably don't care or need to know much about a LM317 linear regulator putting 5V to a terminal block.
For advanced designs, from my and many others experience, the problem escalates to chips having undocumented errata and even datasheets just being wrong. If you don't read the datasheet in the first place in the design, you'll be even more lost trying to figure out what's going on. You ultimately have to reach out to the semiconductor vendor and go back and forth, but if you don't even have that starting knowledge you aren't going to be very productive doing so.
Otherwise if I want a generated design for something simple, I could just use TI's WebBench and the like to generate them, no AI required :shrug:. I still get bit once in awhile with regulators that have compensation loop issues contrary to the datasheet the generator was based on. Hah.
> For simple designs, sure, you probably don't care or need to know much about a LM317 linear regulator putting 5V to a terminal block.
Oddly, this is exactly the type of person for whom AI will result in a failure. You can get AI to generate this bit and instantly know how to integrate it and if it's got a major error/oversight.
I literally tried to do this with a 9V battery and a vape cartridge (simple resistance load) to jerry-rig a vape whose power source had died but the rest worked fine. I had a box of all the most common op-amps, transistors, and some voltage regulators. Also plenty of other passives on-hand, pretty much anything that could be needed. I asked a variety of AI's including ChatGPT-4 and Flux.ai to design this for me. But none of them worked for providing enough current, or even generally the right voltage.
And I didn't know enough about electronics to know what wasn't working.
I use AI a lot in software programming. But I use it for the basic, boring boilerplate stuff that I just don't want to type. It frees up my time and brainpower to spend on the hard parts with trickier data structures and algorithms where there's no "right" answer, just different trade-offs. Using AI for the simple stuff is easy because I can quickly read the code that is output and instantly find any errors before I copy/paste it into my work. But if I ask the AI to make something complicated, it can take awhile for me to see where the errors are, or that there's a subtle structural problem with the architecture that's making things far more complex than they need to be. These parts are better hand-crafted.
>But I use it for the basic, boring boilerplate stuff that I just don't want to type.
The boilerplate isn't stuff you generally need AI for really if you are working professionally in electrical. You just pull-up a premade "design block" your company has standardized on.
One of the big things in electrical is you aren't reinventing the wheel each time in your design once you have products going. You re-use parts especially to get long term economy of scale pricing benefits from part vendors by using the same parts more often. This is for everything from ICs to capacitors.
These design blocks extend so far that not only could you drop in pre-made schematic circuits but also premade layout chunks in the CAD tools.
That's if you don't just open an existing schematic and just rename it to your new schematic and make changes hah
And once your overall design gets boilerplate enough, you just get outsourced. It's why no consumer and industrial power supply module design occurs in the US. That stuff has been outsourced to basically 3 or 4 Taiwanese and Chinese OEMs that just copy and paste the same design into different form factors.
I'm absolutely not an "AI" fanboy/believer, but tasks like parsing design sheet PDFs and producing boilerplate schematics are absolutely within the realm of current LLMs, given enough guardrails.
I've spoken to your team and spent a bit of time using it, I have given this feedback to them as well.
Kicad is a long way from being usable for high speed boards and Flux is currently far less usable than that.
I am sure Flux might be 3-5x more productive for folks slapping an ESP32 and some LEDs down on 2 layers, but it's basically impossible to do any kind of real board with substantial power delivery or high speed routing. I can't even imagine trying to route a 8L board with Flux, much less a 20L board with HDI (can you do HDI?)
Cool beginnings, but ultimately DOA from the difficult fact that I am not going to lock up my business's IP in your proprietary web service. I don't know how you are going to ever get past this particular sticking point: at least with Altium, I have a SchDoc I can parse when things go south - and an offline copy of the software.
As an ex hardware engineer turned software engineer I think what you're aiming for here is great!. I have always thought that a lot of the day to day software development tools would be a huge productivity boost if they were somehow reimagined / re-implimented for hardware design (and probably a ton of other industries, we take the level of tooling for granted).
My old colleagues have started to bring things like bug trackers and SCM to their workflow, but its still ~15year old tech (using SVN for VHDL for example)
I didn't see any mention of versioning, merging changes that kinda of thing. Do you have any plans there?
Are these copilots essentially all the same underlying model, but with various configurations creating a different prompt for the model? And to create your own, you configure it based on your project specs? Just making sure I understand what's happening. If so, it's kind of a cool and "simple" solution to making a model work better. Simple in quotes because nothing in software is actually simple, but the idea is.
But in short: yes, thats how it works as of today! Our docs go into a bit more detail here.
But this space has been moving so fast that usually how it worked last week issn’t how it works this week.
You could easily imagine us leveraging fine tuned world models models on top of prompting in the near future.
Another thing to add is that we use a bunch of different models for our Copilot implementation…some popular off the shelf ones as well as some that we trained ourselves.
The really cool thing that I forgot to mention here is that it makes the model composable
As you provide requirements and specs via properties to your project OR by adding components (which have their own properties) you are implicitly optimizing the model towards your use case (eg Analog Audio Aerospace applications)
These requirements/specs are all in human language…so an added benefit is that you provide more documentation for your human peers too!
We are planning to directly integrate with fabs to place order/get quotes in the near future though