Read the Full Series
This article is one part of a walkthrough detailing how we recreated an NXP i.MX 8M Mini–based computer using Quilter’s physics-driven layout automation.
This PCB autorouter comparison covers the main ways to automate routing in 2026: batch autorouters such as Altium's Situs and Cadence's Allegro PCB Router, guided interactive routers such as Altium's ActiveRoute, the open-source Freerouting for KiCad, reinforcement-learning autorouters such as DeepPCB, LLM agents such as GPT-6 Astra, and Quilter, which places and routes the board together and checks every candidate against physics constraints. Traditional autorouters connect nets on a board you have already placed. Quilter generates complete layout candidates for your team to review. For the history behind that difference, and why completion rate is a poor measure of either, see The PCB Autorouter Was the Right Idea.
This guide is built for engineers and technical decision-makers evaluating automated PCB routing, AI PCB layout, and PCB design automation. You’ll learn how modern autorouters work, how to match tools to your board class and budget, and how to compare them on your own board.
Quilter makes an AI layout tool, so some readers will expect us to be biased. We link each claim about another tool to that tool's own documentation or to a public test. If anything here is inaccurate, let us know and we'll correct it.
What’s Changed in PCB Routing Since 2024?
The biggest change since 2024 is that “automated routing” is no longer a single feature inside your ECAD tool. It’s becoming a distinct workflow: generate candidates, validate them against physics and manufacturing constraints, then choose the best option for final polish.
Traditional autorouters are still rule-driven at their core. You define constraints, set routing rules, and the router tries to connect nets within those rules. On simpler boards this can be helpful. On dense, high-speed, or mixed-signal boards, the classic failure mode hasn’t changed much: you get partial completion, too many vias, awkward detours, and a long tail of manual cleanup.
A traditional router is a geometric solver. It works with widths, clearances, and via rules and has no model of impedance, current capacity, or return paths, so the engineer translates each electrical requirement into a geometric rule by hand. One missing rule can produce a board that passes DRC and misses its impedance target.
AI-powered routing shifts the evaluation from a single routing pass toward exploring multiple layout candidates. Quilter uses physics-driven AI and reinforcement learning to generate candidates and evaluate constraints. Whether this saves time or reduces manual cleanup compared with Altium Designer or KiCad must be established with matched designs, documented settings, and the same acceptance criteria.
AI layout tools differ most in what they are checked against. General language models still fall short on engineering design. In EngDesign, a benchmark of 101 design tasks graded by simulation or evaluation scripts, no model passed more than 35%. On HWE-Bench, a preprint benchmark of 300 board-level schematic design tasks, the best model reached an overall pass rate of 8.15%. Two of the strongest results in AI, AlphaFold and AlphaGo Zero, came from systems trained against objectives they could score directly. Quilter's layout engine is a reinforcement-learning agent trained on boards scored against physics, and that grounding is what "physics-driven" refers to.
If layout is your bottleneck, compare tools on the engineering hours it takes to reach a layout your team will sign off, including setup and cleanup.
How Do Today’s Leading Autorouters Actually Work?
Most “autorouters” fall into four categories: batch autorouting (rule-based or trained with reinforcement learning), guided interactive routing, AI-assisted candidate generation, and LLM agents that operate a CAD tool. They all aim to connect nets, but the way they search for solutions is fundamentally different.
Batch Autorouting: From Rule Engines to Reinforcement Learning
Classic routers use rules and iterative optimization to find routes. Cadence’s Allegro PCB Router lineage (SPECCTRA) is a well-known example of a shape-based router that focuses on connecting signals and resolving conflicts through its routing process. A 2013 guest post on Cadence's PCB blog by Dave Price of DFM describes an approach where the router starts by connecting signals and then resolves constraint conflicts, which is one reason high-end routers can feel more “determined” than entry-level ones. Altium Designer's batch router, Situs, maps the board topologically, defines a routing path first, and then converts it to copper with proven routing algorithms. For the history of Specctra and the routers that followed, see The PCB Autorouter Was the Right Idea.
Strength: strong at enforcing explicit rules at scale.
Tradeoff: the router does not understand intent. It only knows what you told it, and it will happily produce technically legal but practically painful layouts.
A newer generation of batch routers replaces hand-tuned heuristics with reinforcement learning. DeepPCB is one, and in our view it works like an autorouter with a new engine underneath. It offers AI placement as a separate step, and a routing job "never moves parts." Its benchmarks report routing completion and DRC-clean passes, and its help page lists "DDR, length matching and skew, crosstalk, complex BGA fanout, advanced impedance control and RF-specific constraints" among what it doesn't handle yet.
Guided Interactive Routing: Push-and-Shove, Glossing, and Assisted Routing
Modern ECAD tools are excellent at interactive routing: you sketch a path, the router pushes obstacles, respects clearances, and lets you quickly tune. Altium’s ActiveRoute is a good example of “automation that is not a full autorouter.” Altium explicitly describes ActiveRoute as a guided interactive router rather than a complete autorouter, and notes it cannot place vias or handle power routing strategies in the way a full autorouter would.
Analogy: GPS navigation. You are the driver. The tool keeps you from making illegal turns and helps you find a clean path.
AI-Assisted Routing and Layout: Candidate Generation Plus Physics Validation
AI-assisted PCB layout automates more than the connection step. It also handles the reasoning about constraints and physics. The easiest way to think about it is “self-driving” rather than “GPS.” You still set the destination (constraints, priorities, keepouts, impedance targets), but the system generates complete candidates and reports which constraints each candidate met. Quilter compiles the board as one problem: it places and routes together and checks every candidate with Physics Rule Checks: trace width for impedance and current, ground plane under impedance-controlled traces and differential pairs, length match and spacing within each pair, and the distance from decoupling capacitors, crystals, and switching converters to their pins.
LLM Agents That Operate a CAD Tool
General-purpose models now drive CAD tools directly. OpenAI's GPT-6 Astra placed and routed a board in KiCad, and in the public KiCad tests that completed routing, it handed the routing to Freerouting, a conventional autorouter (Eight Public GPT-6 Astra PCB Tests). In Quilter's own test on the 8-layer Project Speedrun SOM (system-on-module), an Astra-led workflow in Altium completed 98.1% of required connections over about 62 hours, and the board still failed the test's acceptance criteria, with 21 open connections, tracks on prohibited ground layers, and failed DDR timing checks (GPT-6 Astra in Altium).
Public tests also show what agents do well. JLCPCB reported that a GPT-6 Astra KiCad project for a 2-layer amplifier passed with 0 ERC and 0 DRC violations under its configured rules, with some rule categories set to be ignored. The routed layout in that project came from an earlier task and was reused unchanged. EDA vendors are building agents that run their existing tools. OpenAI and Synopsys said GPT-Synopsys will directly operate Synopsys tools, which Synopsys calls the ground-truth engineering tools, and Cadence and Siemens check agent output against established EDA engines (Tom's Hardware). For a PCB, an agent's result depends on the router or layout engine it calls.
To evaluate a candidate-based workflow, record each stage: prepare the project, define constraints and design intent, generate candidates, review checks, select a candidate, and export it for final CAD review. Comparing several candidates can help expose tradeoffs, but the number of candidates alone does not establish a speed or quality advantage.
Which Solution Fits Your Budget and Board Complexity?
The right “best autorouter software” choice depends on two things that matter more than brand names:
- Board class: simple vs moderate vs dense/high-speed
- Tolerance for manual cleanup: a little polish vs days of rework
Start from the alternative your team would use otherwise. For many teams that is a layout contractor or service bureau. If test fixtures, variants, and redesigns for obsolete parts wait behind one or two designers, include the contractor quote and the weeks of waiting in the comparison, along with license costs.
Below is a quick self-check that maps common scenarios to tool categories. If you want a single takeaway: traditional tools can be great when you have time and expert operators. AI PCB layout is most valuable when iteration speed and completion confidence are the limiting factors.
A Short Routing-Fit Checklist
- Do you route high-speed constraints like DDR, PCIe, USB, or HDMI?
- Do you routinely spend a full day (or more) cleaning up routing after “automation”?
- Do you need to explore multiple stack-ups, form factors, or placement options quickly?
- Do you want automation that still fits into Altium, Cadence, Siemens, or KiCad workflows?
If you answered yes to any of these, include an AI layout tool in your evaluation.
Budget Tiers in 2026 (Practical, Not Theoretical)
Tier A: Free or near-free (best for learning and simple boards)
- KiCad plus external autorouting via FreeRouting is a common route for budget-conscious teams. FreeRouting describes itself as "an advanced autorouter for all PCB programs that support the standard Specctra or Electra DSN interface," which is why it’s often used as a companion to KiCad workflows. In NextPCB’s public KiCad test, OpenAI’s GPT-6 Astra used the same KiCad and Freerouting setup to finish routing; see what the public GPT-6 Astra PCB tests found. Best fit: 2-layer to 4-layer boards where interactive routing is acceptable and time-to-cleanup is not a critical schedule driver.
Tier B: Prosumer and professional ECAD (best for strong interactive routing)
- Tools like Altium shine when you have designers who can route quickly using interactive features and constraint systems. Assisted routing can reduce effort, but you are still guiding the process. Altium’s own documentation is clear that some automation features are guided interactive routing rather than full autorouting. Best fit: moderate complexity, teams that value a unified environment and strong manual control, and schedules that can absorb polishing time.
Tier C: Enterprise routing stacks (best for sophisticated constraints, complex organizations)
- High-end router ecosystems can do a lot when rules are mature, libraries are clean, and teams know how to drive the tooling. They are often expensive, and the productivity gains depend heavily on process maturity.
Tier D: AI-powered layout systems (best for reducing cycle time and increasing iteration)
- If your actual problem is “layout is the rate limiter,” AI-assisted candidate generation can change the economics. Quilter supports common enterprise and open workflows: its Quickstart documentation lists supported CAD file formats including Altium, KiCad, Cadence Allegro, and Siemens Xpedition. Best fit: boards within Quilter's current capabilities, high iteration pressure, and teams that want to keep engineers focused on architecture, review, and bring-up instead of routing volume.
- What Quilter handles as of November 2026: boards with fewer than 5,000 pins, pin density under 20%, 48 V and 10 A or less, and through-hole vias. Automated fan-out for square-pattern BGAs shipped in May 2026, and impedance calculations are accurate up to 6 GHz. Blind and buried vias and length matching are in development. Quilter is evolving fast, so check What Quilter Does Well and the changelog for the latest.
A Quick Mapping from Board Type to Tool Approach
Simple MCU boards (2-layer)
If you are routing a basic microcontroller board with relaxed constraints, almost any modern tool can get you there. Your deciding factor is usually cost and comfort.
Moderate mixed-signal (4-layer)
Include manual cleanup in the evaluation, rather than relying on an autorouter’s completion percentage. Compare guided interactive routing and candidate-based workflows on the same placement, constraints, and final checks.
Dense high-speed and RF (6 to 12 layers)
For DDR, PCIe, impedance-controlled nets, or RF keepouts, include rework and tuning in the schedule comparison. A complex board is a useful test of constraint handling, but it does not by itself establish which tool will save the most time. Pre-route any section outside Quilter's current capabilities by hand before you submit the board.
What Results Can You Expect from Quilter vs. Traditional Tools?
Measure three outcomes when comparing traditional autorouting with AI-assisted layout: time to an accepted layout, manual cleanup effort, and the number of acceptable candidates available for review.
Completion and cleanup depend on the board, constraints, placement, operator, and acceptance criteria. The table describes evaluation considerations, not measured completion rates or a ranking of cleanup effort.
Side-by-Side Comparison
| Solution | Routing Technology | Typical Completion Rate | Manual Cleanup Effort | Best Fit |
|---|---|---|---|---|
| Quilter | Physics-driven AI, RL-based candidate generation | Measure on the same board and rules | Record all manual cleanup time | High iteration pressure, teams reducing layout as a bottleneck |
| Altium Designer | Interactive routing, push-and-shove, guided assisted routing | Measure on the same board and rules | Record all manual cleanup time | Strong interactive workflows, teams with experienced operators |
| Cadence Allegro (Router ecosystem) | Advanced shape-based routing with extensive constraint systems | Measure on the same board and rules | Record all manual cleanup time | Enterprise environments with mature rules, libraries, and routing specialists |
| DeepPCB | Reinforcement-learning routing in the cloud, a private cloud, or on premises (deeppcb.ai) | Measure on the same board and rules | Record all manual cleanup time | Check DeepPCB's own documentation |
| GPT-6 Astra (LLM agent) | Operates KiCad or Altium; hands routing to the CAD tool's router or Freerouting | Measure on the same board and rules | Record all manual cleanup time | Experiments and scripted CAD tasks |
| KiCad + FreeRouting | Open-source interactive routing plus external rule-based autorouting via DSN | Measure on the same board and rules | Record all manual cleanup time | Budget-sensitive teams, simpler layouts, learning and prototyping |
How to Compare Quilter, Altium, and KiCad Fairly
The layout evaluation guide sets out how to compare layout workflows using matched inputs, explicit acceptance criteria, and recorded human and compute time. It does not provide verified head-to-head results or establish a winner across board classes.
Use these criteria when collecting your own results:
- Measure the whole workflow. Separate compute time from human effort, and include setup, review, correction, and export.
- Record cleanup explicitly. Count manual adjustments and the time needed to meet the same release criteria for every tool.
- Test representative complexity. Include the constraints your team actually uses, repeat runs, and report failures as well as successful results.
- Inspect the checks. Ask what each candidate is checked against, and read the report yourself. In a 2026 IoT Analytics survey of design and engineering decision-makers at 120 manufacturers, 16% trusted AI-supported verification and validation. In the Stack Overflow Developer Survey 2026, the most common answer on trusting AI output, chosen by 48%, was "I trust it when I can easily verify the output." Each candidate comes with a Physics Rule Check report, and candidates with DRC errors aren't surfaced for review.
- Clear the security review first. Ask each vendor where designs are processed and stored, whether self-hosting is available, whether your designs are used for training, and which export-control requirements, such as ITAR, the deployment supports. Get the answers in writing before you commit engineering hours to a trial.
A Concrete Case Example: Project Speedrun (What “Engineering Bandwidth” Looks Like)
In Quilter’s Project Speedrun, Quilter laid out an 843-component, 5,141-pin, two-board Linux computer. Layout services quoted 428 hours; human cleanup took 38.5 hours. The team fully validated the computer with memory, Ethernet, USB, HDMI, and graphics support capable of running hardware-accelerated Chrome and games. Quilter runtime for placement plus routing plus physics validation was reported as 27 hours, with routing completion at 98%.
Engineering judgment stayed with the team; Quilter took the hours of placement and routing.
Here’s How You Can Integrate Quilter Into Your Workflow
The easiest adoption path is to treat Quilter as a layout acceleration layer, not a replacement for your ECAD environment. You keep your existing CAD tools for schematic capture, final DRC, documentation, and fabrication outputs. Quilter sits in the middle where schedule pain usually lives: placement and routing throughput.
Quilter’s Quickstart guidance is explicit about what you need up front: a complete schematic and a supported CAD file format. It lists support for Altium, KiCad, Cadence Allegro, and Siemens Xpedition. From there, the workflow looks like this:
- Prepare your project and constraints
Start with a clean schematic and a linked board file. Define what must not change: board outline, connector placement, keepouts, layer count, and any critical routing intent like diff pairs or impedance targets. - Upload your native project
You upload native design files from the CAD tool your team already uses, including Allegro .brd files and zipped Xpedition project folders, with no intermediate export. The goal is low friction: use the same source of truth you will use for signoff. - Review detected constraints and intent
A major differentiator for AI-assisted layout is that it can surface what it believes the design requires. For example, Quilter detects differential pairs from net classes and net-name suffixes such as +/- or P/N and calculates their trace width and spacing from your stackup. - Generate multiple candidates in parallel
This is where the workflow stops behaving like a classic autorouter. Rather than producing one best-effort result, you get multiple complete candidates to compare, each evaluated against constraints. - Select, polish, and export back to your CAD
Once you pick a candidate, you can run your normal signoff process: DRC, silkscreen cleanup, fab notes, and any final tuning.
Visual Workflow Diagram
Ready to See Faster, Smarter PCB Layout?
If your team is still treating layout as a slow, serial step, 2026 is a good year to change that. Traditional autorouters help on simple boards and noncritical nets. On the boards that fill your layout queue, test an AI layout tool against your current flow. Start by trying Quilter on a real board, compare the candidate outputs to your current flow, and decide based on what matters most: speed, quality, and how quickly you can get to a signed-off design. For next steps, explore the Product pages, Workflow overview, and Support resources on the Quilter site, or request a demo tailored to your board class and constraints.
Frequently Asked Questions
What Is the Best PCB Autorouter?
It depends on the board. For simple boards you already route in Altium, the built-in Situs autorouter and ActiveRoute are the lowest-friction options. KiCad users can add Freerouting at no cost. For the boards that fill your layout queue, test an AI layout tool such as Quilter, which places and routes the board and checks each candidate against physics constraints. Compare all of them on engineering hours to a signed-off layout.
Is There a Free PCB Autorouter?
Yes. Freerouting is open source and works with KiCad and other tools that exchange Specctra DSN files. Quilter also has a free version for non-commercial work.
Which PCB Design Software Has Auto Routing?
Commercial autorouters ship inside Cadence Allegro, Altium Designer, and Siemens Xpedition. KiCad relies on external routers such as Freerouting. Quilter works with native files from all four and returns candidates in the same format.
Is an AI Autorouter the Same as Quilter?
Quilter places and routes together, so it does more than an autorouter, which starts from a finished placement. What is an autorouter? covers the difference in short form.
Can ChatGPT or Claude Lay Out a PCB?
General models can drive a CAD tool and handle small boards. In JLCPCB's test, a GPT-6 Astra KiCad project passed with 0 ERC and 0 DRC violations under the configured rules, though its routed layout was reused from an earlier task, and in NextPCB's test it handed the core routing to Freerouting. On EngDesign, a benchmark of engineering design tasks, the best general model passed about a third. For larger boards with tight constraints, a general model works best as the interface to a layout engine that produces and checks the physical result.
Related Reading
Sources
Altium. "ActiveRoute." Altium Designer Documentation. Accessed November 9, 2026. https://www.altium.com/documentation/altium-designer/pcb/routing/activeroute.
Altium. "Automated Board Layout Using the Situs Topological Autorouter." Altium Designer Documentation. Accessed November 9, 2026. https://www.altium.com/documentation/altium-designer/pcb/routing/situs-topological-autorouter.
Anand, Harsha. "The AI Trust Gap in Design and Engineering Software." IoT Analytics, September 24, 2026. https://iot-analytics.com/the-ai-trust-gap-in-design-and-engineering-software/.
DeepPCB. "DeepPCB." Accessed November 9, 2026. https://deeppcb.ai/.
DeepPCB. "Benchmarks." Accessed November 10, 2026. https://deeppcb.ai/benchmarks/.
DeepPCB. "Help." Accessed November 10, 2026. https://deeppcb.ai/help/.
Downing, Shane. "The State of Agentic AI in Chip Design Tools in 2026: Cadence, Synopsys, and Siemens All Pitch Autonomous Engineers." Tom's Hardware, September 30, 2026. https://www.tomshardware.com/tech-industry/semiconductors/the-state-of-agentic-ai-in-chip-design-tools-in-2026-cadence-synopsys-and-siemens-all-pitch-autonomous-engineers.
Freerouting. "Freerouting." GitHub repository. Accessed November 10, 2026. https://github.com/freerouting/freerouting.
Guo, et al. "Toward Engineering AGI: Benchmarking the Engineering Design Capabilities of LLMs." arXiv:2509.16204, 2025. https://arxiv.org/abs/2509.16204.
JLCPCB. "GPT-6 Astra PCB Design in KiCad: From Brief to Verified Native Project." September 14, 2026. https://jlcpcb.com/blog/gpt-6-astra-pcb-design-in-kicad.
Jumper, John, et al. "Highly Accurate Protein Structure Prediction with AlphaFold." Nature 596 (2021): 583–89. https://www.nature.com/articles/s41586-021-03819-2.
Price, Dave. "How Do You Route Your Highly Constrained PCBs? (Part 1 of 2)." System, PCB, & Package Design (Cadence Community blog), August 30, 2013. https://community.cadence.com/cadence_blogs_8/b/pcb/posts/how-do-you-route-your-highly-constrained-pcbs-part-1-of-2.
Qiu, Weibo, Yinhao Xiao, and Runyu Pan. "HWE-Bench: Can Language Models Perform Board-level Schematic Designs?" arXiv:2603.18102, March 18, 2026. https://arxiv.org/abs/2603.18102.
Quilter. "Changelog." Accessed November 10, 2026. https://www.quilter.ai/changelog.
Quilter. "Downloading Candidates." Quilter Documentation. Accessed November 10, 2026. https://docs.quilter.ai/downloading-candidates.
Quilter. "From Freerouting to DRC: Eight Public GPT-6 Astra PCB Tests." Quilter Blog, October 27, 2026. https://www.quilter.ai/blog/gpt-6-astra-pcb-test-results.
Quilter. "GPT-6 Astra in Altium: 62 Hours on an 8-Layer PCB." Quilter Blog, October 30, 2026. https://www.quilter.ai/blog/gpt-6-astra-altium-pcb-routing-test.
Quilter. "PCB Layout Efficiency Across Board Types: A Reproducible Comparison." Quilter Blog, November 17, 2025, updated September 25, 2026. https://www.quilter.ai/blog/a-data-driven-look-at-pcb-layout-efficiency-in-2025-quilter-vs-altium-designer-vs-kicad.
Quilter. "Physics Rule Checks: Overview." Quilter Documentation. Accessed November 10, 2026. https://docs.quilter.ai/physics-rule-checks-prcs/overview.
Quilter. "Project Speedrun: The First Computer Designed by Physics-Driven AI." Accessed November 10, 2026. https://www.quilter.ai/project-speedrun.
Quilter. "Quickstart." Quilter Documentation. Accessed November 10, 2026. https://docs.quilter.ai/get-started/quickstart.
Quilter. "What Quilter Does Well." Quilter Documentation. Accessed November 10, 2026. https://docs.quilter.ai/about-quilter/what-quilter-does-well.
Siemens. "Auto-Routing Methodologies and Best Practices Webinar." Siemens Support Center, KB000112900. Accessed November 10, 2026. https://support.sw.siemens.com/en-US/okba/KB000112900_EN_US/Auto-Routing-Methodologies-and-Best-Practices-webinar/index.html.
Silver, David, et al. "Mastering the Game of Go without Human Knowledge." Nature 550 (2017): 354–59. https://www.nature.com/articles/nature24270.
Stack Overflow. "Trust in AI Tools." Stack Overflow Developer Survey 2026. Accessed November 9, 2026. https://survey.stackoverflow.co/2026/ai/data/ai-trust.
Synopsys. "OpenAI and Synopsys Announce GPT-Synopsys: Frontier Intelligence to Revolutionize Chip Design." Press release, September 30, 2026. https://investor.synopsys.com/news/news-details/2026/OpenAI-and-Synopsys-Announce-GPT-Synopsys-Frontier-Intelligence-to-Revolutionize-Chip-Design/default.aspx.
Zheng, Carmen. "Can GPT-6 Astra Automate PCB Routing? Don’t Panic, AI Isn’t Replacing PCB Designers Yet." NextPCB, September 2026. https://www.nextpcb.com/blog/gpt-6-astra-kicad-pcb-routing-test.













.webp)



















