Six autorouters, eight boards, one referee

Between 14 and 16 October 2026 we ran six autorouters on eight professionally routed open-hardware boards, on one workstation, with one time cap, and let KiCad score every board that came back. The full table, the method and the list of what we do not claim are on the benchmarks page. This post is the short version and what stood out.

The setup

• Eight boards, 2 to 4 layers, 125 to 997 connections. A KiCad demo, three Raspberry Pi reference designs including the CM5 IO board, three OLIMEX SoM carriers and the Antmicro scalenode CM4 baseboard. Each is used under its upstream licence and linked to its source files.
• Placement only. Every router starts from the human placement with tracks, vias and copper pours removed. Every net, ground included, is the router’s to connect.
• Same rules, same cap, same machine. One rule floor per board, published and applied to every tool and to the referee. 1200 seconds of wall clock per board. One workstation.
• KiCad is the referee. Completion is KiCad’s own ratsnest, read with kicad-cli: 100 × (1 − unconnected / connections). DRC errors are the ones a tool added over the human board’s own findings. No tool scores itself.
• The six. Koppar, Freerouting 2.5.0, tscircuit 0.0.979, KiCad Routing Tools 0.1.9, TraceMaker 0.9.0 and the hosted Zap! service, each run as shipped with its defaults.

The result

Koppar fully routed all eight boards with zero DRC errors added, in both of its modes: the default, which keeps every via out of pin pads so the board needs no special fab process, and the filled-via mode for a board ordered with filled and capped vias.

Board	Layers	Connections	Koppar	Closest other result
pic-programmer	2	125	100%, 0 DRC, 8 vias, 6 s	Zap! 100%, 0 DRC, 28 vias, 56 s
rp2040-minimal	2	150	100%, 0 DRC, 116 vias, 5 s	TraceMaker 100%, 0 DRC, 53 vias, 20 s
pico-vga	4	180	100%, 0 DRC, 96 vias, 8 s	Zap! 100%, 0 DRC, 83 vias, 17 s
cm5io	4	431	100%, 0 DRC, 877 vias, 887 s	TraceMaker 94.9%, 0 DRC, 412 vias, 1203 s
rk3328-som-evb	4	462	100%, 0 DRC, 359 vias, 46 s	TraceMaker 100%, 0 DRC, 272 vias, 1202 s
scalenode-cm4	4	446	100%, 0 DRC, 428 vias, 79 s	TraceMaker 100%, 0 DRC, 224 vias, 832 s
stmp157-som-evb	4	905	100%, 0 DRC, 1131 vias, 484 s	TraceMaker 100%, 0 DRC, 856 vias, 1203 s
imx8mp-som-evb	4	997	100%, 0 DRC, 1429 vias, 1194 s	TraceMaker 99.3%, 0 DRC, 795 vias, 1202 s

Koppar’s time is the whole product path, everything a user’s own board goes through. Times for the other tools are their own wall clock on the same machine.

What stood out

• The CM5 IO board separated the field. 431 connections on four layers. Koppar finished it. TraceMaker reached 94.9% at the cap. Freerouting and Zap! timed out, tscircuit returned no board, and KiCad Routing Tools reached 50% with 88 DRC errors.
• Completion and DRC are different questions. tscircuit reached 100% on five boards, more than any other router besides Koppar, but with 115 to 1052 DRC errors added per board. Freerouting reached 100% on the two easiest boards with 5 and 8 errors. A board that is fully connected and fails DRC is not a board a fab will take.
• TraceMaker was the closest on quality. It added zero DRC errors on every board it returned and drilled fewer vias than Koppar on every 4-layer board. It fully routed four of the eight and needed most of the 20-minute cap on the carriers.
• Zap! was fast on the small boards and clean. Zero DRC errors on every board it returned, 100% on the three smallest, 96 to 97% on two carriers, two timeouts, and one board refused because the human design has a footprint outside the outline.
• The two-layer boards hide the differences. Five of six routers fully routed pic-programmer. The four-layer carriers are where the field spreads out, which is why the next run has more of them.

What we do not claim

• No human parity. Koppar routes every board here, but the human layout is still better on vias, track length and plane integrity. Completion and DRC do not score planes.
• Each tool was run as shipped. A tool’s author may know settings that do better. The boards, rules, harnesses and raw results are public in autorouter-bench, so anyone can rerun them.
• Completion is not quality. Rail sizing, decoupling distance and return paths are not scored yet.
• Eight boards is a small sample. 25 more public open-hardware boards, 2 to 8 layers, are pinned in the repository with their sources and checksums. The run on them is coming.

Run it on your own board

The same router is online at tools.koppar.ai: upload a placed .kicad_pcb with its .kicad_pro and get it back routed with a KiCad DRC report. A board up to 300 connections needs no account.