Rhaamo kicad amiga2000 amiga2000

pcb-corpus

235 real KiCad boards from 170 open-source hardware repositories, prepared as a benchmark for PCB placement and routing tools. Each board comes as its designer shipped it and as an unrouted copy, with its own project rules.

It exists because the public routing benchmarks are built on KiCad 5 files with no project file, so every board is judged against KiCad's default rules. These are current designs judged against their own.

What is in a board folder

manifest.json lists every board with its statistics. splits.json is the original three-way split by size: easy (under 40 connections, 37 boards), medium (40 to 149, 112) and hard (150 and up, 86).

What qualified

A board is here if:

• its repository carries an open-hardware or permissive licence that GitHub recognises,
• a .kicad_pro sits beside the .kicad_pcb,
• it opens in KiCad 9, has 1 to 4 copper layers and at least 5 parts on nets,
• and its designer's routing leaves no connection open.

158 of the 235 boards have error-level DRC violations under their own rules as designed. ref_drv in the manifest is that count. For a "clean pass" comparison (fully connected, zero DRC errors), use the 77 boards where it is 0.

Mix: 178 two-layer, 57 four-layer. Licences: MIT 75, GPL-3.0 67, CERN-OHL-S-2.0 25, CERN-OHL-P-2.0 25, Apache-2.0 19, CC-BY-SA-4.0 10, CERN-OHL-W-2.0 5, CC-BY-4.0 4, BSD-3-Clause 3, CC0-1.0 2.

Groups and difficulty

features.json describes every board: size and density, stackup, packages, project rules, how its designer routed it, wiring estimates, and what is on it. groups.json sorts the boards into overlapping groups, and GROUPS.md has the full table with a baseline score for each group.

The "on the board" and controller groups are detected from net, footprint and value names. They are useful for slicing results, not ground truth; expect some mislabels.

Difficulty is measured. Each board's designer placement is routed by a baseline router and judged by KiCad DRC. The first baseline is TraceMaker with a 30-second budget (baselines/tracemaker-30s.json):

Over all 235 boards that baseline routes 20% clean and connects 94.5% of connections. On the 77 clean-reference boards it routes 49% clean. What separates easy from hard, by share routed clean:

• Easiest: untangled 45%, single-sided-routing 43%, small 43%, no-pour 38%, through-hole 35%.
• Hardest: xlarge, fpga and bga 0%, tangled 3%, crystal 6%, large 8%, double-sided 9%, fine-pitch 10%, four-layer 11%.

The tiers will move as baselines are added: a board is solved once any baseline routes it clean. Add one with python3 tools/baseline.py --name <name> --cmd '<router> {in} -o {out}', then python3 tools/groups.py.

Using it

• Routing: route unrouted.kicad_pcb, refill the zones, and run kicad-cli pcb drc --severity-error on the result with board.kicad_pro copied beside it.
• Placement: move the parts of unrouted.kicad_pcb, then route and judge the same way. The designer's placement is the reference.
• A router that reads the board file directly will treat a stale zone fill as copper in its way. The unrouted boards have their fills cleared for that reason; refill after routing, before DRC.

The collector that built this, and harnesses that run it, are in punkfab/circuit-skills under evals/corpus.

Licences

This repository is a collection. Each board keeps the licence of its source repository, copied into its folder; see ATTRIBUTION.md for the full list. Several are share-alike (GPL-3.0, CERN-OHL-S, CC-BY-SA): if you redistribute or modify those boards, their terms apply to you. unrouted.kicad_pcb is a modified copy (routing removed) under the same licence as its board.

The licence recorded for a board is the one GitHub detects for its repository. If you are an author and a board here is under different terms than recorded, or you want it removed, open an issue and it will be taken down.

The manifest, splits, README and attribution files are released under CC0-1.0.