How Long Does It Take to Route a PCB? Data From 55,000+ Completed Routing Jobs
It depends on the board, and anyone who answers with one number is measuring the wrong thing.
Ask how long it takes to route a PCB and you usually get a single number back. That number is close to meaningless, for two reasons.
The first is that routing time scales with the board. A 60-airwire two-layer board and a 1,200-airwire eight-layer board are not the same job, and quoting one median across both tells you nothing about yours.
The second is that router time and end-to-end turnaround are different measurements. Router time is how long the router takes to produce routing output on a board you submit. End-to-end turnaround is how long it takes to get from a board that needs routing to a board an engineer is willing to send to fabrication. The gap between them is human work, and it is most of the clock.
This article answers both, by board size, and keeps them apart.
The dataset: 21,000+ engineers have created 100,000+ board designs on the DeepPCB platform. 55,000+ of those went through a completed routing job, meaning the router ran and produced output. Not uploads. Not drafts. Completed routing runs.
Router Time, by Board Size
This is the table that actually answers the question. Every figure is a median across completed routing jobs in that band, on the DeepPCB production platform.
| Airwires on the board | Median router time | 90th percentile |
|---|---|---|
| Under 100 | 2 minutes | 10 minutes |
| 100 to 299 | 7 minutes | 31 minutes |
| 300 to 599 | 9 minutes | 45 minutes |
| 600 and above | 16 minutes | 60 minutes |
Find your airwire count, read across. If your board has 400 airwires, the 9-minute row is your number.
For completeness, the all-boards figures are a median of about 2.5 minutes, a mean of about 6.6 minutes, a 90th percentile of about 15 minutes and a 95th percentile of about 27 minutes. We publish them because they are the honest aggregate, not because they will predict your board.
What Drives the Time
Airwire count, which dominates everything else
More connections means a larger search space, and it grows faster than linearly because every connection interacts with the ones near it. The size table above is really a picture of this one variable.
Two named boards from our published benchmarks bracket the range. The STRF RF mixed-signal transceiver (STM32L432 with an nRF24L01P, 98 airwires, four layers, hand-routed RF front end protected) routed to 100%, DRC clean, in 2 minutes 06 seconds. At the other end, a 5,208-airwire eight-layer board reached 100% completion in roughly 14 hours. Same router. The board is the whole difference.
Density rather than dimensions
Pins per unit area predicts difficulty better than raw board size. A congested small board can take longer than a roomy large one with the same connection count.
Constraint complexity
Differential pairs, protected regions, per-net-class rules and clearance hierarchies all narrow what counts as a legal solution, and narrower is slower. A four-layer board with 88 components, 43 nets and one differential pair routed 210 of 210 connections at 100%, DRC clean, with 42 vias, in 18 minutes 08 seconds. That is far above the platform median on a board most engineers would call ordinary, and the constraints are the reason.
End-to-End Turnaround: a 2-to-3-Day Board in About 3 Hours
The public anchor for the whole job comes from Alain-Sam Cohen on the record in EEJournal, “2,000+ AI-Routed PCBs a Week”, published 28 July 2026: a board an experienced designer routes by hand in two to three days completes in about three hours, plus 30 to 60 minutes of engineer cleanup.
That is not the same measurement as the router times above, and it is not a slower version of them:
| Measure | What it counts | Figure |
|---|---|---|
| Router time | One router run on one submitted board | 2 minutes at under 100 airwires, rising with the board |
| End-to-end turnaround | The whole job: setup, review, re-runs, human finishing | About 3 hours plus 30 to 60 minutes cleanup, against 2 to 3 days by hand |
The three-hour figure is a different board from the named examples elsewhere on this page, walked through end to end rather than timed as a single run.
Both are true. Only the second one is what your schedule cares about.
Where the Other Hours Go
The gap between the two answers is almost entirely human time, and naming it is more useful than hiding it:
- Import and board setup. Net classes, layer stackup, plane definitions, which components and pre-routed sections to protect. Get this wrong and the router optimises against the wrong problem quickly instead of slowly.
- Constraint capture. Historically the largest single sink, and the reason a fast engine did not produce fast projects. Cooper, the assistant that captures constraints in natural language and fixes board setup, shipped publicly on 18 June 2026 to attack exactly this step.
- Review and re-run. At minutes per attempt this is cheap. At hours per attempt it costs you a day.
- Finishing. The last few connections, plus anything you would rather own by hand.
Compared With Routing It By Hand
Adoption is no longer the interesting question. In the Weidmuller USA 2026 PCB Design Tools and Component Selection Survey (fielded by EETech Research, 400 North American engineers, April to May 2026, released 31 July 2026), 91% of engineers said they had used AI-based tools in their PCB design workflow, while 75% viewed AI as a productivity and acceleration layer within existing workflows rather than an autonomous step.
That second figure is the more useful one, and it happens to describe how this actually works in practice. The engine is not a replacement for the engineer on the clock. It removes the mechanical bulk so the engineer spends their hours on the parts that need judgement.
Which is why the honest comparison is the end-to-end one. About three hours plus cleanup against two to three days by hand is a real saving. Two minutes against two to three days is a slogan.
And speed only counts next to completion. Across those 55,000+ completed routing jobs: 96.1% mean completion, 100% median completion, and 87% of boards reach 95%+ completion, DRC clean by default. Completion here means connectivity, the fraction of target connections routed. It is not a measure of design quality, and a high completion figure does not tell you the layout is one you would sign off on.
What These Numbers Do Not Tell You
- Router time is not a guarantee. It is a distribution measured after the fact. Your board sits somewhere in it, and we cannot tell you where before we see it.
- The aggregate median is not your board. Use the size band that matches your design.
- Completion is connectivity, never quality. 96.1% mean completion says connections were made. It says nothing about whether the result meets your review standard.
- These times cover routing only. Component placement stays your job.
- The tail is real. Above a few hundred airwires the 90th percentile runs to tens of minutes, and the largest boards run for hours. If you have a hard deadline on a dense board, plan around the tail.
- Not capabilities. DDR routing, length matching, signal integrity analysis, EM and thermal analysis are outside what the engine does. Do not read a fast routing time as coverage of any of them.
- We deliver routed boards, not reports. The output is a board file you open in your own tool.
Data and Methodology
Dataset. All platform figures come from the DeepPCB production platform, not a curated benchmark set. Population: boards with a completed routing job, meaning the engine ran and produced output. Uploads, drafts and abandoned boards are excluded.
Sample. 55,000+ completed routing jobs across 100,000+ board designs created by 21,000+ engineers. The size table is computed on the subset of those jobs carrying a clean routing-time measurement, about 30,000 boards.
Routing time basis. Successful revision processing time over completed routing jobs, and per-revision processing time on the largest boards, where three separate timing fields agree to within a few minutes.
Completion basis. Connected airwires divided by total airwires on the latest revision. Connectivity only.
Named board figures. Individual reference boards are single runs, quoted as measured, and are not statistical claims. The STRF and Smart MCU boards are on our benchmarks page, and more worked examples are coming soon.
End-to-end turnaround. Not a platform measurement. It is a public statement in EEJournal, 28 July 2026, describing a full-board workflow including human finishing.
As of end of July 2026.
Related Reading
- PCB Autorouter Benchmark 2026, the full completion and via distributions across the platform
- DeepPCB vs Quilter on three open-source boards, public boards and public tools, so the comparison is reproducible
- Published benchmarks, the named reference boards
- Alain-Sam Cohen on the AI in Business podcast on applying reinforcement learning to placement and routing
Try it on a real board: app.deeppcb.ai. Specctra/DSN, JSON, KiCad, Altium, Zuken and EasyEDA files supported. A board under 100 airwires typically routes in about two minutes; larger boards take longer, and the whole job including setup and finishing takes longer still.
FAQ
Q: How long does it take to route a PCB with AI?
It depends on the size of the board. On the DeepPCB platform, boards under 100 airwires have a median router time of about 2 minutes, boards of 100 to 299 airwires about 7 minutes, and boards of 300 to 599 airwires about 9 minutes, measured across completed routing jobs. The all-boards median is about 2.5 minutes, but a single figure will not predict a larger board, so use the band that matches your design. Separately, end-to-end turnaround including board setup, review and human finishing is much longer: a board an experienced designer routes by hand in two to three days completes in about three hours plus 30 to 60 minutes of cleanup.
Q: Is AI routing faster than manual PCB routing?
It is faster on a board that has been set up well, but “faster” is the wrong frame. What it really does is accelerate the process rather than replace it. The engine takes the mechanical bulk of the connections, the engineer stays in charge of when to intervene and which nets to own by hand, and the finishing pass is still human. This matches how the market sees it: in a 2026 survey of 400 North American engineers, 75% described AI as a productivity and acceleration layer within their existing workflow rather than an autonomous step. The practical claim is not that the AI beats an engineer. It is that an engineer with it gets more done than the same engineer without it, because the tedious majority of the board stops consuming their hours.
Q: Why do some PCBs take hours to route instead of minutes?
Mostly airwire count, then density and constraints. A 60-airwire two-layer board and a 1,200-airwire board are different problems, and the search space grows faster than the connection count. Tight constraints such as differential pairs, protected regions and per-net-class clearance rules narrow the set of legal solutions further, which is why a 210-connection board with a differential pair took 18 minutes while a 98-airwire RF board took 2 minutes. At the extreme, a 5,208-airwire eight-layer board reached 100% completion in roughly 14 hours.
