Autorouting High Speed Digital PCBs: Why It Falls Short

Autorouters have improved enormously, and for many boards they are genuinely useful. Yet on a high speed digital design they still produce results that experienced engineers reject, and the reason is not that the algorithms are weak. It is that the router optimises for the objective it was given, which is usually completion and clearance, while high speed routing is judged against electrical behaviour that is difficult to express as a rule and impossible to infer from a netlist.

What Autorouters Do Well

A modern router is excellent at the problem it was designed for: connecting a large number of nets through a defined space without violating geometric rules. It explores thousands of routing options in seconds, handles dense fanout areas efficiently and never tires. On a board with hundreds of low speed control signals, it can produce a routing that a human would take a day to draw and would probably draw worse.

That capability is worth using. The failure mode is not using an autorouter at all, but using it on the nets where the electrical behaviour matters and then accepting the result because the design rule check passes. Rules describe what is geometrically legal, and legality is not the same as adequacy.

Constraints the Router Cannot Invent

A router can obey a constraint it is given, and it cannot invent one. Impedance targets, length tolerances, maximum stub lengths, allowed layer transitions and forbidden regions all have to be stated as explicit routing constraints before the router starts. Where a designer sets up a board with default rules and presses autoroute, the tool will happily produce a routing that crosses reference plane splits, changes layers without return vias and matches lengths by accident.

Almost all of the effort in high speed routing is therefore front loaded. Our notes on high speed PCB layout floorplanning describe the decisions that have to be made before a single trace is drawn, and those decisions are exactly the ones an autorouter cannot make for you.

Reference Plane Continuity

Reference plane continuity is the clearest example. A router sees a plane as an obstacle to route around, and a split in that plane as an opportunity to pass through. Electrically, a fast signal crossing a split forces its return current to take a long detour, which raises radiation and crosstalk and degrades the edge. Encoding plane continuity as a mechanical rule is possible but fragile, and verifying it afterwards is easier than constraining it in advance.

Layer transitions have the same character. Every via that changes layers needs a return via nearby, and the placement of that return via depends on which planes the signal traverses. A human can see the structure; a rule set can only approximate it.

Length Matching and Timing Groups

Length matching is a set of relationships rather than a set of numbers. Which traces belong to the same group, which one is the reference for the others, how much skew the interface tolerates and whether the matching has to hold differentially as well as single ended are all design decisions. A router can match lengths inside a group it has been told about, and it cannot decide that two nets belong together.

Matching also interacts with everything else. Serpentine sections added late consume routing space, create crosstalk between their own turns and sometimes push a trace across a plane boundary. A human routing those nets usually allocates the space for matching during floorplanning, which is why the result looks deliberate rather than patched.

Differential Pairs and Impedance Control

Differential pairs are the case where automated routing most often fails visibly. Keeping a pair coupled, maintaining the correct intra pair spacing, avoiding asymmetric detours and providing a continuous reference are all things a router can attempt and rarely achieves across a long route. The result is usually a pair that is electrically mismatched even though both traces have the same length.

Impedance control adds another layer of difficulty, because the target depends on the stackup and on the layer. A trace that changes layer changes its impedance environment, and unless the tool models that change and compensates, the finished board will show discontinuities that no geometric check reveals. Our notes on microstrip and stripline routing describe how the two environments differ.

Where Automation Still Helps

The productive division of labour is to route the critical nets manually and let the tool handle the rest. Power distribution, low speed control signals, general purpose input output and the fanout of large arrays are all areas where automation saves time without compromising the electrical design. Even there, the result should be reviewed rather than accepted.

A second productive use is exploration. Running an autorouter on a new floorplan shows whether the routing channels are wide enough to complete the connections, which is valuable feedback during planning rather than after the layout is finished. Treating the output as a feasibility study rather than as a finished layout changes how it should be judged.

Reviewing an Autorouted Board

If an autoroute is used for any part of a high speed design, the review has to be more thorough rather than less. Check that no fast net crosses a plane split, that every layer transition has a return via, that pairs stayed coupled, that length groups meet their tolerances and that the impedance target is the same at both ends of each route.

This is the same set of checks that applies to a hand routed board, which is the point. The tool changes how the copper got there; it does not change what the copper has to do. The structure of a proper PCB design review process makes that clear, and it applies whether the routing was drawn by a person or generated by software.

What It Costs to Route by Hand

The honest counter argument is time. Routing a dense high speed board by hand takes days, and the schedule pressure to press autoroute is real. The costs of the shortcut, however, arrive later and in worse places: an EMC test failure, an intermittent link that only fails at temperature, or a redesign after the product has been tooled. Spending the time on the nets that matter, and letting the tool do the rest, is usually both faster overall and cheaper than repairing an automated result.

A useful compromise is to hand route one representative net of each class first, verify the geometry against the constraints, and then use that route as the template for the rest of the class. The tool copies the discipline, the human sets it, and the review burden stays manageable.

FAQ

Is autorouting ever acceptable on a high speed board? Yes, for the non critical portions, and as a feasibility tool during planning. Using it on impedance controlled or length matched nets usually creates more review work than routing those nets by hand, because the resulting geometry has to be checked against constraints the tool was never given.

Can I encode all the constraints and then trust the router? You can encode a great deal, and modern tools support impedance, length and layer constraints well. The limits are the relational and contextual decisions: which nets belong in a group, where a plane split is acceptable, and which discontinuities the link can tolerate. Those remain engineering judgements.

What is the biggest risk of an autorouted high speed design? A design that passes every automated check and still fails, because the checks describe geometry and the failures are electrical. The second largest risk is a layout that is very difficult to modify later, since automated routing tends to fill space in ways that leave no room for the change that always comes.