Help

It is young, and what is in it works. Importing, the copper and thermal solve, DC and transient simulation, the report and the 3D view all work today, and we use them ourselves.

In practice, some things are not switched on yet (the AI features and the paid plans), and you will meet rough edges we have not found. Your account and anything we owe you, such as the waitlist credit, carry over unchanged when we leave beta. Nothing gets reset.

If something is wrong, tell us. During the beta, that helps us more than at any other time.

No, so don't sign a board off on its word alone. Amptrace is a second pair of eyes for what is easy to miss: a track too thin for the current it carries, a rail that sags under load, a part that runs hot. It can only check what it models, so a board with no findings can still have faults.

Keep everything you already do: your own review, your EDA tool's DRC and ERC, a review in your team, and a prototype on the bench. Amptrace is one more check beside them.

That depends on the model behind each number. The copper is exact, because it is read from your files, not redrawn. The parts are simplified: a switching regulator is solved as an averaged element instead of switched cycle by cycle, and a value the AI estimated without a datasheet is marked as an estimate.

Read a result as “this is in the right range, and this is where to look”, not as a measurement. Where a margin is thin, measure it.

No. Those check your design against its rules: clearances, unconnected pins, a schematic and board that disagree. Amptrace asks what happens electrically and thermally once current flows. Some of its checks overlap with theirs but do not replace them, so run both.

Yes. The aim is a first prototype that works, not no prototype at all. How much a board radiates, noise that depends on the layout, assembly faults, the firmware, and anything a model leaves out only show up on real hardware under a real load.

No, and none of them is ever run. The files are read directly: a custom s-expression parser for KiCad's .kicad_pcb, .kicad_pro and .kicad_sch, Altium's binary compound file for copper, netlist and values, and Eagle's XML with the deviceset's pin-to-pad mapping applied so a sheet net resolves to a board net.

There is no plugin or export step, and no version to match: you open the project files you already have.

Yes. RS-274X/X2 plus Excellon drill, and the BOM and pick-and-place files if you have them. Nets are reconstructed from exact copper overlap, pads from the solder-mask openings, and part positions from the CPL. Gerbers are the copper, so the copper is as real as in any other import.

What a gerber set cannot say is what each part is. That comes from the BOM, or from telling the assistant the part number yourself. Without either, the board is geometry with unknown parts on it, and the solve says so instead of guessing.

Then the netlist is enough. A board is synthesised from the schematic so the circuit can be simulated: parts, pins and connectivity come across, and the DC and transient solves run on that.

It cannot give you anything that needs a layout's copper: current density per track, IR drop, trace heating. There is no artwork to measure, so those checks are skipped, not estimated. Import the board later and the same circuit gains them.

Import and simulation never upload it; they run in your browser or on your machine. AI modelling sends only what the model needs to answer: part numbers, pin names, connected nets, and any datasheet you attached. The Data & Privacy panel logs every request with the exact payload it carried, and on every plan, the free trial included, you can have it hold each request for your approval before it is sent. You can also tell the assistant to send nothing about your board and still ask it questions.

Attach its datasheet and the model is built from that document instead of from the model's general knowledge. The Inspector has the button, next to the part it is modelling.

With no datasheet the panel marks the values as the AI's best estimate, and they can be badly wrong without looking wrong: a 1.2 mΩ MOSFET came back as 30 mΩ, a factor of twenty-five out.

In the desktop app datasheets fetch themselves. In a browser most manufacturers refuse the request, so there you attach the PDF by hand (the last question on this page explains why).

Open the demo board and everything that acts on it: pan and zoom the copper, browse nets, place sources, run the DC solve, run a transient, watch the heatmap animate. That is the whole simulator, on a board we provide. You need an account to bring your own design or to use the AI features.

AI modelling stops until the allowance resets, and tells you so. Your boards, the copper solver and every analysis keep working, because none of them run on our machines.

What spends it: each part you ask the AI to model. A part it recognises from its number is cheap; one whose datasheet has to be found and read costs several times as much, because the datasheet goes to the model with the question. A part answered out of the shared cache costs nothing. The figure on each card is therefore roughly how many parts a month it covers, and your account page shows what you have used.

Everything else is unmetered. The solver, the report, the 3D view and every analysis run on your own machine, however many boards you put through them.

The part that does the work, yes. Import, the copper and thermal solve, every analysis, the report and the 3D view all run on your own machine and need no connection, which is also why none of them is metered.

AI modelling needs the service, because the model is not on your machine and never will be, and so do the checks that decide which plan you are on. On a flight you can solve a board and read the results, but you can't ask the assistant about a part it has not seen.

One installation of the desktop app, which is part of Pro. Reinstalling on the same machine does not use a second seat, and you can deactivate one from your account page at any time to free it up.

Two things, both limits of a browser tab. Your projects become real files, with native Open and Save As, a document you can double-click and a recent-files list, so a design can live in git or Dropbox instead of in one browser's storage.

And datasheets fetch themselves. Manufacturers serve most datasheet PDFs without the header a browser requires, so a web page is refused them; we saw that for five parts out of five. A model built without the datasheet falls back on the AI's general knowledge, which can be badly wrong. The installed app fetches from your own address and gets them.

We tried a server-side proxy for the browser and measured it against 150 real datasheet URLs. Half failed, all for reasons that come from one address fetching for everybody: hotlink refusals, rate limits, bot timeouts. In the browser the app still tries, since the occasional vendor allows it, and otherwise asks you to attach the PDF yourself.