ProjekinAja. Embedded Systems Stage 5 · PCB Design Lesson 5.6
● Stage 5 · Lesson 5.6

Routing: Traces, Current, and Return Paths

Copper is a resistor and a loop antenna. Route as if both were true, because both are.

⏱ ~27 min read 🧠 IPC-2221 and loop area 🛠 KiCad net classes and DRC

🎯 By the end you'll understand

How to size a trace from current and allowable temperature rise using IPC-2221's relation, why the same trace carries half the current on an inner layer, and why the answer that standard gives is a survival minimum rather than a design width. How to compute resistance and voltage drop in squares of copper, and how much current a via really carries. What a return path is doing under every trace you draw, what a split in the reference plane does to it, and how to price that in loop area and decibels of radiated field. The routing rules for a differential pair that matter and the ones that are folklore. And Beacon Rev A's routing plan, including the charge trace that passes IPC-2221 and still ruins a humidity reading.

KEY INSIGHTEvery trace is doing two jobs, and only one of them is in the netlist

The charge trace was six mils wide. IPC-2221 said it was fine at 500 mA, and IPC-2221 was right — it never gets hot enough to fail. It just runs six degrees warm, three millimetres from a humidity sensor with a ±3% requirement.

A netlist says which pads must be connected. It says nothing about the two properties that routing actually decides. The first is resistance: copper is a conductor, not a wire from a schematic, and a trace carrying current drops voltage and turns some of it into heat right where the trace happens to run. The second is loop area: every trace and its return enclose a region, and that region is a small antenna in both directions — it radiates what your circuit does and it receives what the world does.

Neither property appears in any tool's default checks. DRC checks clearances against rules you typed in. It has no opinion about whether 500 mA through a thin trace will warm a sensor, or whether the return current for a signal you just routed over a plane split has anywhere to go. The most useful mental shift in this lesson is to stop seeing traces as connections and start seeing them as shapes with current in them — because that is what the physics sees.

🔑 Hold onto this

Two questions, asked of every net before you route it. How much current, and where does the heat land? And where does the return current flow, and what is the area of the loop? If you can answer both for every net on a board, the routing is essentially finished before you start.

CORE CONCEPTWidth, current, and heat

The standard everyone quotes is IPC-2221, whose current-carrying relation is:

📖 IPC-2221 conductor sizing

I = k · ΔT0.44 · A0.725

with I in amperes, ΔT the allowed temperature rise in °C, A the conductor's cross-sectional area in square mils, and k = 0.048 for external (outer-layer) conductors and 0.024 for internal ones. Cross-section is width × copper thickness: 1 oz copper is 1.378 mil (35 µm), 0.5 oz is 0.689 mil, 2 oz is 2.756 mil.

The halved k for internal layers is the standard's blunt statement that a buried conductor cannot convect heat away — it must conduct it out through the laminate. Same width, same copper, half the current for the same rise.

Worked: Beacon's 500 mA charge path

R6 requires USB-C charging at 500 mA, so the path from the MCP73831's VBAT pin to the JST connector carries half an amp. Take a 10 °C allowable rise on an outer layer in 1 oz copper.

🧪 Worked: minimum width on an outer layer

Rearranged: A = ( I / (k · ΔT0.44) )1/0.725.
ΔT0.44 = 100.44 = 2.754, so k·ΔT0.44 = 0.048 × 2.754 = 0.1322.
A = (0.5 / 0.1322)1.379 = 3.7821.379 = 6.27 mil².
Width = 6.27 / 1.378 = 4.55 mil = 0.115 mm.

🧪 Worked: the same current on an inner layer, 0.5 oz

k halves to 0.024, so A = (0.5 / 0.0661)1.379 = 7.5641.379 = 16.3 mil².
Inner copper on Beacon's stackup is 0.5 oz = 0.689 mil, so width = 16.3 / 0.689 = 23.7 mil = 0.60 mm.

Five times wider for the same current and the same temperature rise, from two changes that both live in the stackup rather than in your routing. This is why power goes on the outer layers whenever it can, and why "just take it through on an inner layer" is a decision with a number attached.

Why 0.115 mm is the wrong answer anyway

Here is the failure the lesson opens with. Suppose you route the charge path at 6 mil (0.152 mm) — comfortably above the IPC minimum, and a natural default width. Invert the relation to find the actual rise: A = 6 × 1.378 = 8.27 mil², A0.725 = 4.626, so

ΔT = ( 0.5 / (0.048 × 4.626) )1/0.44 = 2.2522.273 ≈ 6.3 °C

Six degrees. IPC-2221 is satisfied — the trace will not fail, will not delaminate, will not fuse. And six degrees of warm copper running past the SHT40 is, by Lesson 5.5's arithmetic, roughly eighteen points of relative-humidity error against a ±3% requirement. The standard answers "will this trace survive", not "will this board work". Those are different questions and only one of them is yours to answer.

⚠ IPC-2221's charts are old, and IPC-2152 replaced them

The curves behind that equation come from measurements on isolated traces in still air with no plane nearby, published decades ago. Real boards are not that: an adjacent ground plane is an enormous heat spreader, board thickness matters, copper elsewhere matters. IPC-2152 (2009) is the current standard and accounts for those, and its answers differ — often permitting substantially more current for a given rise when a plane is present, and less in still, thermally isolated situations. IPC-2221's equation survives because it is a closed form anyone can compute in one line. Use it as a fast sanity check with margin; for anything thermally critical use IPC-2152's charts, a thermal simulation, or a measurement with a thermal camera on the first board.

Resistance, in squares

The other half of copper's day job. Sheet resistance makes this arithmetic trivial: 1 oz copper is about 0.49 mΩ per square, and a "square" is any length of trace equal to its width. So a trace's resistance is 0.49 mΩ × (length ÷ width), with both in the same units — no unit conversions, no resistivity lookup.

🧪 Worked: Beacon's charge path at two widths

The path is about 20 mm long. At 0.5 mm wide that is 40 squares: R = 40 × 0.49 mΩ = 19.6 mΩ. At 500 mA the drop is 9.8 mV and the dissipation 4.9 mW, spread over 20 mm of copper — negligible.

At the IPC minimum of 0.115 mm it is 174 squares: R = 85 mΩ, a 43 mV drop and 21 mW concentrated in a thin strip. The volts do not matter much (the charger's constant-current phase does not care, and by the constant-voltage phase the current has tapered), but the milliwatts are in exactly the wrong place. Halve the copper, quadruple the heat density.

Copper's resistance also rises about 0.393% per °C, so a warm trace is a slightly worse trace — a small effect here, a compounding one in high-current designs.

Via current

A via's conductor is its plated barrel: circumference × plating thickness. IPC-6012 sets the minimum copper in the hole at about 20 µm for Class 2 and 25 µm (1 mil) for Class 3; fabs commonly plate to the 25 µm figure, which is what the arithmetic below assumes. For a 0.3 mm finished hole that is π × 11.8 mil × 1 mil ≈ 37 mil² of cross-section, which the same relation turns into roughly 1.8 A at a 10 °C rise. Treat that as an upper bound and use two or three vias on any power connection anyway — not for the current, but because parallel vias cut the inductance of the layer transition, which matters far more than the amperes.

SEE ITTrace and return-path lab

Two experiments on the same idea. The first sizes copper against current and shows what the standard actually promises. The second takes a routed signal and removes the ground under it, which is the single most common way a board that passes every check radiates.

⚙ Interactive · Trace and return-path lab

Deterministic. Temperature rise inverts IPC-2221's closed form; resistance uses sheet resistance in squares. The return-path model is an order-of-magnitude geometric argument, labelled where it is — the mechanism is exact, the millimetres are illustrative.

0.50 A
0.15 mm
20 mm
20 mm
6.0 mm
—
—
—
—
—
—
—
—

In the first mode, set the current to 0.5 A and walk the width down from 0.5 mm. Notice how flat the curve is until it is not: temperature rise goes as roughly the inverse of width to the 1.65 power, so the last tenth of a millimetre costs more than the first half. Then switch the layer to inner and watch the whole curve lift. In the second mode, start on solid ground and then introduce the slot: the trace is unchanged, the netlist is unchanged, DRC is unchanged, and the loop area — the thing that actually radiates — grows by an order of magnitude.

CORE CONCEPTReturn paths and what a plane split does to them

From Lesson 5.4: above the tens-of-kilohertz region, return current takes the path of least inductance, which means it flows in the reference plane directly under the trace, spreading only a millimetre or so either side. The loop enclosed by trace and return is therefore length × dielectric height: on Beacon's stackup, 0.21 mm² per millimetre of trace. That is a very small loop, which is the whole point of paying for four layers.

Now cut the plane. A slot for a connector's through-holes, a row of vias whose antipads merge into a line, the boundary between two rails on the power plane — any of these is a barrier the return current cannot cross. The return does not stop; it goes around. It travels along the barrier to the nearest place it can cross, comes back on the other side, and rejoins under the trace.

The same trace, twice — the difference is underneath it Intact reference plane return hugs the trace · loop = length × 0.21 mm ≈ 4.2 mm² for a 20 mm trace A slot under the trace slot detour d extra enclosed area of order d² ≈ 40 mm² at d = 6 mm — 20 dB more field Nothing about the trace changed. No net is disconnected, no clearance is violated, and no design-rule check has an opinion.
The "order of d²" argument is a hand calculation, not a field solve: the return is displaced by roughly the distance to the nearest crossing, so the extra enclosed area scales with that distance squared. Take the decibels as a magnitude, not a measurement. What is exact is the direction and the reason.

Two consequences follow, and both are routing rules rather than review findings.

Pours and stitching are the same idea from the other side. A ground pour on a signal layer is only useful if it is genuinely tied to the plane — an unstitched pour is a floating piece of metal that couples noise rather than draining it. Stitch it, and stitch it at a spacing set by the highest frequency you care about: at 2.45 GHz in FR-4 the wavelength is about 58 mm, so λ/20 is 2.9 mm. Near the module, stitch every 3 mm or tighter; elsewhere on Beacon, every 5 mm is plenty.

IN PRACTICEPairs, pours, and the rules worth arguing about

Differential pairs: what matters and in what order

  1. A continuous reference under the whole pair. This is first by a long way. A split under a differential pair converts differential signalling into common-mode current, which is what actually radiates and what actually fails an emissions scan.
  2. Constant spacing. The pair's differential impedance depends on the gap; a gap that opens up around an obstacle is an impedance discontinuity. Route them together, turn them together, and if one must dodge something, make both dodge.
  3. Symmetric layer transitions. If the pair changes layer, both traces change at the same place with matching via geometry, and a ground via goes next to them.
  4. Length matching. Last, and much less important than people assume at low speeds. What matters is intra-pair skew, and Beacon's full-speed USB has an enormous skew budget. Match within 0.1 mm anyway: it is free, the router does it for you, and the habit transfers to boards where it is the whole ball game.
  5. Distance from other nets. The "3W" guideline — keep other traces three trace-widths away edge-to-edge — is a rule of thumb about how fast crosstalk decays with separation, not a standard. Treat it as a default to justify departing from, not a law.

Teardrops, acid traps, and the folklore audit

Two rules that appear in every checklist deserve unequal respect.

Teardrops — the fillet where a trace meets a pad or via — earn their place. They add copper exactly where drill wander could otherwise break out of a small annular ring, and they distribute mechanical stress at the joint on a board that flexes. KiCad 8 and 9 generate them across the board in one operation. Turn them on.

Acid traps — acute angles where etchant was supposed to pool and over-etch — are largely a legacy concern. Modern spray etching does not leave standing chemistry in a corner, and most fabs will tell you so if you ask. The rule persists in DRC checklists mainly because it is easy to check. Routing at 45° rather than in acute angles is still good practice for other reasons: fields are less concentrated at gentle bends and the board is easier to read. Do it because it is tidy, not because you are afraid of acid.

The one to actually worry about in the same family is a slot or sliver in the solder mask between closely spaced pads, which the fab will simply remove if it is below their minimum. That one is Lesson 5.8's problem and it is real.

BEACONBeacon Rev A's routing plan

Routing starts with net classes, because a net class is a rule you only have to think about once. Beacon's classes go into Board Setup before a single track:

Net classNetsWidthWhy
DefaultEverything not listed below0.20 mmComfortably above the fab's 0.09 mm minimum, so etch variation never matters
Power+3V3, VSYS0.40 mmCarries the radio burst's ~40 mA average and its peaks; wide is also low-inductance
ChargeVBUS, VBAT, the MCP73831 output0.60 mm500 mA under R6. IPC-2221 asks for 0.115 mm; this is five times that, chosen for heat, not survival
USBD− / D+ (GPIO18 / GPIO19)0.24 mm, 0.13 mm gapA coupled pair over solid ground; not impedance-controlled (Lesson 5.4) but routed as though it were
I2CSDA / SCL (GPIO6 / GPIO7)0.20 mm400 kHz, slow; free to live on L4 and cross whatever it likes
1

The hot loop first, before any other track

Cin's pads to the buck's VIN and GND, on the top layer, with the return on the same layer where possible and immediately adjacent vias where not. This loop is drawn once and then locked. Everything else routes around it.

2

The charge path, wide and short, away from the sensor

USB-C VBUS to the MCP73831, its output to the protection pair and the JST, all at 0.60 mm on the top layer, staying in the board's lower-right corner. At 0.60 mm the temperature rise at 500 mA is well under a degree, and — more to the point — the path never comes within 15 mm of the SHT40. This is the failure from the opening quotation, designed out by geometry rather than by width alone.

3

The USB pair, top layer, straight, over uncut ground

GPIO18 and GPIO19, about 15 mm from receptacle to module. Constant 0.13 mm gap, 22 Ω series resistors placed symmetrically so both legs see the same geometry, TVS on the connector side, no layer changes at all — the easiest way to keep a pair's reference continuous is never to leave the layer that has one.

4

Decoupling loops, then the rest of the top layer

Each capacitor's ground via placed before its trace, so the loop is short by construction rather than by luck.

5

Slow signals to L4

I²C to both sensors on GPIO6/GPIO7, the WS2812B data line on GPIO5, the button on GPIO9, the accelerometer’s INT1 on GPIO3, the Tag-Connect fan-out and the pogo pads. Every one of these may cross the splits on L3 without consequence, which is the entire justification for the SIG/GND/PWR/SIG stackup.

6

Pours, stitching, and a look at layer 2

Ground pour on L1 and L4, stitched at 5 mm generally and 3 mm around the module. Then the step people skip: display layer 2 alone, at full zoom, and confirm it is a single uninterrupted region. Every connector's pin row, every via cluster, every antipad is a candidate slot.

beacon.kicad_proexcerpt · net classes as JSON
{
  "net_settings": {
    "classes": [
      { "name": "Default", "track_width": 0.20, "clearance": 0.15,
        "via_diameter": 0.60, "via_drill": 0.30 },
      { "name": "Charge", "track_width": 0.60, "clearance": 0.20,
        "via_diameter": 0.80, "via_drill": 0.40 },
      { "name": "USB", "track_width": 0.24, "diff_pair_gap": 0.13,
        "diff_pair_width": 0.24, "clearance": 0.20 }
    ]
  }
}
; Charge is 5x the IPC-2221 minimum. The reason is written on the schematic sheet:
; "0.60 mm chosen for temperature rise near U4 (SHT40), not for conductor survival."

That comment is the point of the whole lesson. A width without a reason is a width somebody will "optimise" later to reclaim board area, and they will be able to prove it still passes IPC-2221 while they do it.

WEAKNESSWhere routing bites

✗ Sizing traces from a calculator and calling it engineering

Online trace-width calculators implement IPC-2221's equation and return a number to three decimal places, which reads as precision and is not. The equation ignores adjacent copper, planes, board thickness and airflow, and it was fitted to measurements that IPC itself replaced with IPC-2152. Worse, the number answers a question about conductor survival that is almost never the binding constraint on a small sensor board. Use it as a floor, then choose the real width from where the heat lands and how much voltage you can afford to lose.

✗ Routing on an inner layer without checking which k applies

Two things change when a power net dives to layer 3. The standard halves the allowable current for a given rise, and Beacon's inner layers are 0.5 oz rather than 1 oz — half the copper again. A width that carried 500 mA comfortably on the top layer needs to be five times wider inside. This bites most often when a router auto-routes a power net through an inner layer to get past an obstacle and nobody notices the layer change, because the connection is correct and DRC is silent.

✗ Believing the ground pour is a plane

A copper pour on a signal layer looks exactly like a plane in the editor. It is not one: it is whatever copper survived after the traces on that layer took their space. On Beacon's L4 the pour is generous in some regions and reduced to isthmuses in others, and the isthmuses are where a return current would have to squeeze through — if any return current depended on that layer, which is exactly why nothing fast lives there. The check is visual and takes thirty seconds: view each layer alone and ask whether the copper you are relying on is actually continuous between the two points that need it.

KEY INSIGHTHow this connects

The copper is drawn. Next: the one region of the board where the right amount of copper is none at all.

GLOSSARYKey terms introduced

IPC-2221 / IPC-2152
Conductor-sizing standards. IPC-2221 gives the closed form I = k·ΔT⁰·⁴⁴·A⁰·⁷²⁵ from older isolated-trace data; IPC-2152 supersedes it with chart-based results that account for planes, board thickness and environment.
Sheet resistance
Resistance of one square of a conductive layer, independent of the square's size. About 0.49 mΩ per square for 1 oz copper, so R = 0.49 mΩ × length ÷ width.
Return path
The conductor carrying current back from a load to its source. Above the tens-of-kilohertz region it flows in the reference plane directly beneath the trace, because that minimises loop inductance.
Plane split
Any barrier in the reference plane — a slot, a rail boundary, a row of merged antipads — that a return current cannot cross, forcing a detour and enlarging the loop.
Bridging capacitor
A capacitor placed at the point where a signal crosses a split between two rails, giving the return current an AC path across. Only works where the crossing is.
Stitching
Vias tying pours on different layers to the plane, spaced at or below a twentieth of the shortest wavelength of concern (about 2.9 mm at 2.45 GHz in FR-4).
Teardrop
A copper fillet where a trace joins a pad or via; adds tolerance against drill breakout and reduces mechanical stress concentration.
Intra-pair skew
The length or delay difference between the two legs of a differential pair; converts differential signal into common mode, which is the part that radiates.

ACTIVE RECALLCheck your understanding

Answer in your own words first, then reveal.

1A 6-mil outer-layer trace in 1 oz copper carries Beacon's 500 mA charge current. Show it satisfies IPC-2221 at a 10 °C budget, then explain why it is still the wrong trace.›

Cross-section is 6 mil × 1.378 mil = 8.27 mil², and 8.270.725 = 4.626, so the current allowed at ΔT = 10 °C is 0.048 × 100.44 × 4.626 = 0.61 A — comfortably above 500 mA. Inverting for the actual rise at 500 mA gives ΔT = (0.5 / (0.048 × 4.626))1/0.44 ≈ 6.3 °C. The trace survives. It is still wrong because six degrees of warm copper is a heat source, and if it passes near the SHT40 that turns into roughly three points of relative-humidity error per degree — around eighteen points against a ±3% requirement. IPC-2221 answers whether the conductor survives, not whether the product meets its specification.

2The same net has to cross the board on an inner layer instead. What happens to the required width, and why twice over?›

Two independent penalties compound. First, IPC-2221 halves k from 0.048 to 0.024 for internal conductors, because a buried trace cannot convect heat away — that alone requires about 2.6× the cross-sectional area for the same rise. Second, Beacon's inner layers are 0.5 oz rather than the outer 1 oz, halving the thickness for a given width. Together, the 0.115 mm that suffices on the outer layer becomes about 0.60 mm inside — roughly five times wider. The practical rule: keep power on the outer layers, and if a power net changes layer, re-check its width rather than assuming the net class still applies.

3A signal is routed 20 mm over a solid plane, then rerouted identically over a plane with a slot in it. Nothing in DRC changes. What changed physically, and what are the two legitimate fixes?›

The return current can no longer flow directly beneath the trace. It travels along the barrier to the nearest crossing point and back, so the enclosed loop grows from roughly length × dielectric height (about 4.2 mm² on this stackup) to that plus an area of order the detour distance squared — an order of magnitude larger for a six-millimetre detour, which is around 20 dB more radiated field at the same current and frequency. The two fixes: move the trace so it does not cross the barrier, which is always the better answer; or provide a crossing exactly where the trace crosses — a pair of stitching vias if the barrier separates two ground regions, a bridging capacitor if it separates two rails. The fix only works at the crossing point; a capacitor bridging the same two rails elsewhere on the board does nothing for this loop.

4Rank these four differential-pair rules by how much they matter for Beacon's USB, and justify the ranking: length matching, continuous reference, constant gap, 3W spacing from other nets.›

Continuous reference is first and not close. A split under the pair converts differential current into common-mode current, and common mode is what radiates and what fails an emissions scan. Constant gap is second: the differential impedance depends on the spacing, so a gap that opens around an obstacle is a discontinuity — though on a 15 mm full-speed run this is a discipline issue rather than a functional one. 3W spacing is third and is a rule of thumb about crosstalk decay rather than a standard; it is a sensible default to depart from with a reason. Length matching is last for this board: full-speed USB has a very large skew budget and a 15 mm run cannot consume it. Match to 0.1 mm anyway because it is free and the habit transfers to designs where skew is the binding constraint.

Next: the one region where the right amount of copper is none →