Decoupling, Grounding, and Return Paths
Current does not flow to ground. It flows back to where it came from, by the path of least impedance — which is not the path you drew.
🎯 By the end you'll understand
Why a capacitor sits next to every power pin, what it is actually supplying, and why the distance between it and the pin is measured in nanohenries rather than millimetres. What a capacitor's self-resonant frequency is and why the "100 nF plus a bulk cap" rule exists — along with the honest argument about whether the traditional decade-spread of values is a good idea at all. Where return current physically flows, why it hugs the trace above a few kilohertz, and what happens when a slot in a plane makes it go the long way around. Ground bounce, loop area, and the reason a four-layer board with a solid ground plane is not a luxury on an RF design. And Beacon's own power distribution, laid out with numbers.
KEY INSIGHTThe schematic shows connections. The board decides impedances.
On a schematic, every wire named GND is the same node and every supply rail is an ideal source. On a board, "GND" is a sheet of copper with current flowing through it in loops you did not draw, and the rail is a source at the far end of a wire that fights every change in current. Decoupling is the practice of making the schematic's fiction locally true.
A colleague brings you a board that resets. Not always — only when the RGB LED turns on, and only sometimes then. The firmware is innocent: the reset happens before any of their code runs after the LED write. The supply looks perfect on a multimeter, 3.30 V steady. On a scope with a short ground spring instead of the long clip lead, the same rail shows a 700 mV notch lasting a couple of microseconds at exactly the moment the LED lights. The board's schematic has a 100 nF capacitor across the module's supply, drawn right next to it. On the actual board that capacitor is 15 mm away, on the other side of a connector, with its ground via nowhere near the module's. The schematic was right and the board was wrong, and nothing in the netlist could tell the difference.
Two equations carry this entire lesson. ΔV = ΔI·Δt / C — a capacitor sags when it supplies charge, and how much depends on how long before something else takes over. V = L·di/dt — a wire opposes change in current, and how hard depends on the loop's inductance, which is set by geometry, not by the net name. Every decoupling rule, every ground-plane argument, and every "it works on the bench but not in the field" story is one of these two with real numbers in it.
CORE CONCEPTThe capacitor as a local battery with a very short shift
A chip does not draw constant current. It draws current in steps: a clock edge switches millions of gates at once, a radio's power amplifier keys on, an LED driver turns on a channel. Each step is a demand for charge that arrives faster than the regulator, several centimetres away and with a control loop measured in microseconds, can respond to. Something local has to cover the gap. That is the entire job of a decoupling capacitor: it is a charge reservoir positioned close enough to the pin that it can answer before the pin notices anyone is missing.
How big does it need to be? Directly from ΔV = ΔI·Δt/C. Suppose a load steps by 60 mA and the local capacitor must cover it for 200 ns before the bulk capacitor's current has ramped up. Allow a 50 mV droop:
C ≥ ΔI·Δt / ΔV = (60 mA × 200 ns) / 50 mV = 12 nC / 50 mV = 240 nF. So a 100 nF alone would droop 120 mV over that window, and a 1 µF would droop 12 mV. Now change the assumption: if the cap must cover the load for 5 µs — which is what happens when there is no bulk capacitor and the regulator's loop is the next tier — the requirement becomes 60 mA × 5 µs / 50 mV = 6 µF. Same load, same droop target, 25× the capacitance, purely because the next tier down was further away in time. This is why decoupling is described as a hierarchy: each capacitor only has to cover the interval until the next, larger, slower one takes over.
That is one axis. The second axis is speed, and it is where the arithmetic stops being about capacitance at all.
HOW IT WORKSCopper is inductance, and inductance is the real specification
Every loop of current has inductance. The rule of thumb worth carrying around: a narrow trace has roughly 1 nH per millimetre when its return path is far away, and closer to 0.5–0.6 nH per millimetre when it runs directly over a ground plane a few tenths of a millimetre below. A via through a 1.6 mm board is a few tenths of a nanohenry. A capacitor's own package contributes its equivalent series inductance (ESL): roughly 0.7 nH for an 0402, around 1 nH for an 0805, more for anything larger or with longer terminations. These are engineering rules of thumb, not datasheet guarantees, but they are accurate enough to make every decision in this lesson.
Now put a number on the opening story. The 100 nF capacitor 15 mm from the pin, with its ground returning by a route rather than through a plane beneath it, sits in a loop of maybe 68 nH once you count the trace out, the return, the vias, and the package. A clock edge that steps 50 mA in 2 ns produces
V = L · di/dt. With L = 68 nH and di/dt = 50 mA / 2 ns = 2.5 × 107 A/s: V = 68 × 10−9 × 2.5 × 107 = 1.7 V. The capacitor has plenty of charge. It simply cannot deliver it fast enough through that much inductance. Move the same capacitor to 2 mm over a plane and the loop is about 3 nH, giving 75 mV. Nothing changed but geometry, and the failure went away.
The frequency-domain version of the same fact is the one printed in every application note: a real capacitor is a series RLC, so its impedance falls with frequency until the inductance takes over, and the turning point is its self-resonant frequency:
fSRF = 1 / (2π√(L·C)), where L is the ESL plus the mounting and trace inductance. Below fSRF the part behaves as a capacitor (|Z| = 1/2πfC, falling). At fSRF the impedance is just the ESR, a few tens of milliohms. Above fSRF it behaves as an inductor (|Z| = 2πfL, rising) and is no longer decoupling anything. A 10 µF 0805 with about 1.2 nH of total loop resonates near 1.5 MHz; a 100 nF 0402 resonates near 19 MHz on its own ESL alone, and near 9 MHz once you add the 2.4 nH of vias and 2 mm of copper it is actually mounted in; the same 100 nF placed 15 mm away in a 68 nH loop, near 1.9 MHz — which is to say, placing it badly turned your fast capacitor into a slow one.
This is why the classic rule is 100 nF at every pin plus one bulk capacitor. The bulk part holds the charge for the slow, big events; the small part has low enough inductance to respond to the fast ones. Together their impedance curves overlap and the rail stays low-impedance across a wide band.
The traditional advice goes further and recommends a decade spread — 10 µF, 100 nF, 10 nF, 1 nF — on the theory that each covers the frequency band above the last. A substantial body of power-integrity work disagrees, for a concrete reason: two capacitors of different values in parallel form a resonant circuit with each other, and between their self-resonant frequencies there is an anti-resonance where the parallel impedance is higher than either part alone. A carelessly chosen decade ladder can produce an impedance peak exactly where the chip is noisiest. The competing recommendation is to use several capacitors of the same value in parallel, which lowers both the impedance and the effective inductance by roughly the count, without creating peaks between them.
Who is right depends on how fast your board is. For a 160 MHz RISC-V microcontroller and a BLE radio on a 35 × 25 mm board, the honest answer is that either approach works and the placement matters far more than the value ladder. For a 1 GHz processor with a controlled-impedance power delivery network, it is a simulation problem, not a rule-of-thumb problem. Be suspicious of anyone who states either position without asking about your edge rates.
DEFINITIONReturn current: where it actually goes
Current flows in loops. Always. There is no exception, no matter how many nets are named GND. When a chip pulls 50 mA from its supply pin, that same 50 mA leaves through its ground pin and travels back to the source. The only question is by which route — and the route is chosen by physics, not by the designer.
At DC, return current takes the path of least resistance
It spreads out across the whole plane, fanning between the load's ground pin and the source's, because copper resistance is what it is minimising. Nothing surprising here, and it is the mental model most people keep.
Above a few kilohertz, it takes the path of least inductance
And least inductance means smallest loop area, which means the return crowds into the copper directly beneath the signal trace, mirroring its path. The crossover between the two behaviours typically falls somewhere between a few kilohertz and about a hundred kilohertz for ordinary board geometries — which is to say every digital edge, every clock, and every switching regulator on your board is firmly in the second regime.
So the plane under a trace is part of that trace's circuit
Cut the plane and the return current cannot follow. It detours around the cut, and the loop area — hence the inductance, hence the voltage generated by V = L·di/dt, hence the radiated field — grows by the size of the detour. A 20 mm detour on a signal that would otherwise have had a 0.2 mm loop height is not a small change; it is two orders of magnitude in enclosed area.
And shared return paths couple circuits that share nothing else
If two circuits' return currents flow through the same piece of copper, the voltage that one develops across that copper's inductance appears in series with the other's ground reference. This is common-impedance coupling, and it is how a motor driver injects noise into an ADC that is on the other side of the board and connected to nothing.
Ground bounce is the same effect seen from inside a chip. When several outputs switch simultaneously, their combined return current flows through the inductance of the package's ground bond and the board's ground connection. Take eight outputs each switching 20 mA in 2 ns through a shared 5 nH: V = 5 nH × (0.16 A / 2 ns) = 0.4 V. For those two nanoseconds the chip's internal ground is 0.4 V above the board's, every input threshold moves with it, and an input sitting near its switching point can register a phantom edge. This is why the fix for "my button generates two interrupts when the LEDs change" is sometimes a decoupling capacitor rather than a debounce routine.
SEE ITThe decoupling bench: place a capacitor, watch the rail
An ESP32-C3-MINI-1 running from 3.3 V. Choose a load step, pick a decoupling capacitor and drag it away from the power pin, decide whether there is a bulk capacitor and what the return path looks like. The widget computes the loop inductance, the resulting self-resonant frequency, and the actual voltage at the module's supply pin during the event. Two lines matter: 3.0 V is the module's minimum operating voltage from its datasheet — below it you are out of specification even if nothing visibly breaks — and 2.6 V is where Beacon's brown-out detector is set, below which the chip resets and your firmware never finds out why.
Three things to try. One: with everything healthy, drag the distance from 2 mm to 15 mm on the core-clock-edge event and watch the inductive notch alone push the rail through both thresholds — the capacitor's value never changed and its charge was never the problem. Two: select the LED event and remove the bulk capacitor. A single 100 nF, even mounted perfectly, has to cover 60 mA for five microseconds until the regulator responds, which is 3 V of sag it does not have. That is the failure in the opening story, and the fix is a 10 µF, not a better 100 nF. Three: keep everything good but change the return path to "no plane". The capacitor is still 2 mm from the pin. The board still fails, because the return current has to travel 40 mm to get home and the loop is what sets the inductance — which is the entire argument for the ground plane in the middle of Beacon's four-layer stack.
IN PRACTICEPlanes, splits, stitching, and the death of star grounding
Star grounding — every circuit's ground brought back by its own dedicated conductor to a single point — is genuinely correct advice, for audio and instrumentation at frequencies where return current still follows the path of least resistance. Applied to a board with megahertz edges, it is actively harmful: it forces return currents onto long, defined routes, guaranteeing large loops precisely where you least want them. On any multilayer board carrying digital signals, an unbroken ground plane wins, and it wins for a reason worth stating precisely: a plane lets every return current find its own minimum-inductance path without you having to know in advance what that path is. It is the only ground topology that solves a problem you have not analysed yet.
| Practice | What it does | When it is right | When it bites |
|---|---|---|---|
| Solid ground plane | Gives every return current a low-inductance path directly under its trace | Essentially always on 4+ layers, and mandatory under an RF module | Only when it is not actually solid — see the next two rows |
| Routing a trace across a plane split | Forces the return current to detour around the gap, enlarging the loop | Never on a fast signal | The detour adds tens of nanohenries and turns the trace into an efficient antenna. This is one of the most common findings in an EMC pre-scan (Lesson 6.4). |
| Slots for connectors or large pads | Same effect, created accidentally by a row of through-hole pins or a keep-out | Unavoidable sometimes — route around, not across | You will not see it in the schematic and DRC will not flag it. It is caught by looking at the plane layer with the signals visible. |
| Stitching vias along a return path | Ties the ground layers together so return current can follow a signal that changes layer | Beside every via that carries a fast signal between layers | Missing them means the return has to find the nearest ground via, which may be centimetres away — the loop is real even though the schematic shows none |
| Separate "analog" and "digital" ground islands | Physically separates return currents, joined at one point | Mixed-signal boards where a converter's datasheet asks for it and you can keep both circuits entirely over their own island | Any trace crossing the boundary now has no return under it. More boards are ruined by well-intentioned ground splits than saved by them; the usual better answer is one plane and careful placement. |
When you place a decoupling capacitor, do not think "put it near the pin". Think: draw the loop. From the capacitor's positive terminal, along the copper, into the pin; out of the ground pin, into its via, through the plane, back up the capacitor's ground via. Trace it with a finger on the layout. The area that finger encloses is the inductance. That habit catches every failure in this lesson — the far capacitor, the shared via, the slot, the missing plane — and it is the same habit Lesson 5.6 will formalise as return-path routing.
BEACONBeacon's power distribution, with numbers
Beacon's board is 35 × 25 mm and four layers, in the canon stackup: signal on top, a solid ground plane on layer 2, a power layer on layer 3, signal on the bottom. That layer-2 plane is 0.1 mm or so below the top layer in a standard 1.6 mm four-layer stack, which is what makes the "0.5 nH per millimetre" figure apply to every top-layer trace on the board rather than the 1 nH of a plane-less design. It is not decoration; it is the reason the rest of these numbers work.
| Where | What | Why that, there |
|---|---|---|
| ESP32-C3-MINI-1 3V3 pin | 100 nF (0402) within ~2 mm, own ground via beside its pad | Covers the module's fast internal switching. Total loop ≈ 3 nH, self-resonance around 9 MHz, |Z| about 3 Ω at a 175 MHz knee. This is the capacitor that stops the core's edges from bouncing the rail. |
| Same pin, second tier | 10 µF (0805) within ~5 mm | The module's radio keys on and off in bursts of tens of milliamps lasting hundreds of microseconds — far too slow for the 100 nF's charge and far too fast for the regulator. This is the part that stops the LED-and-radio class of failure. |
| Regulator output | 10 µF, per the regulator's datasheet | Stability, not decoupling. A buck converter's output capacitance is part of its control loop, and the datasheet's value and ESR range are a requirement rather than a suggestion. This is chosen properly in Lesson 4.2. |
| SHT40 and LIS2DH12 | 100 nF each, at the part | Both are low-current, but both switch internally and both are on a shared I2C bus whose edges are also load steps. The cost is two 0402s and the benefit is that neither sensor's supply moves when the other converts. |
| WS2812B-2020 | 100 nF on the switched LED rail, at the LED | The addressable LED sinks current in PWM bursts as it drives its three channels. Its capacitor sits downstream of the high-side P-MOSFET that gates the LED off in sleep (Lesson 2.7), which means it is discharged in sleep and must be charged at every turn-on — an inrush event, handled in Lesson 2.6. |
| Module antenna region | Ground plane pulled back under the antenna, module at the board edge | Here the plane's job flips: it is the antenna's reference on one side and must be absent under the radiating element. Copper where the antenna needs air detunes it. This is Lesson 5.7's subject and the one place on Beacon where "more ground plane" is wrong. |
Add up the cost of all this decoupling: five 0402 capacitors and two 0805s, perhaps IDR 1,500 in parts at 25-unit quantities against R5's IDR 150,000 unit budget — around 1 %. It consumes no current in sleep beyond the leakage of ceramic dielectric, which for X7R parts at these values is nanoamps and does not register against the canon's 8 µA sleep floor. It costs board area on a 35 × 25 mm board, which is the real constraint, and it costs placement discipline in Lesson 5.5. There is no version of this trade where you win by leaving capacitors off.
Schematic v0 gains a power-distribution block with the decoupling drawn next to the pin it serves rather than gathered in a corner of the sheet — a schematic convention that exists purely to communicate placement intent to whoever does the layout, which on this project is you in Stage 5. It also gains two notes that are constraints rather than components: "layer 2 is an unbroken ground plane; no signal crosses a plane split", and "every decoupling capacitor's ground via within 1 mm of its pad". Those two sentences carry more of Beacon's reliability than any component value on the sheet.
WEAKNESSWhere this bites
The standard scope probe's ground lead is a 10–15 cm wire, forming a loop of a hundred nanohenries or more with the probe tip. That loop is a low-pass filter and an antenna at the same time: it hides the fast notch you are hunting and simultaneously picks up whatever is switching nearby, so you see a clean trace with some unrelated ringing on it. A 700 mV, 50 ns notch simply does not appear. Use a ground spring or a short pigtail soldered to the ground pad next to the point you are probing, and probe at the chip's pins, not at the connector. Lesson 1.7 introduced this; here it is the difference between finding the bug and rewriting firmware for a week.
When a rail misbehaves, the reflex is to add parts: another 100 nF, then a 10 nF, then a 1 nF, scattered wherever there is room. If the problem is inductance — and on a fast edge it usually is — a capacitor placed 12 mm away contributes almost nothing no matter how many of it you fit, because the loop it sits in is the dominant term and paralleling long loops barely helps. Meanwhile each added value creates a new anti-resonance with its neighbours. The first move is always to shorten the loop of the capacitor you already have: closer pad, its own ground via, no shared via with the chip. If that is genuinely impossible, then more capacitance, and preferably more of the same value.
A rail that sags to 2.9 V for 400 ns on a 3.3 V module is out of the datasheet's specification and works anyway, most of the time, on most units, at room temperature. Then the pilot ships to a warmer site, the battery ages so its internal resistance rises, the regulator's dropout grows, and a margin that was never measured becomes a field failure rate that nobody can reproduce on a desk. Margin is not a luxury you add if there is time; it is the difference between a design and a coincidence. When the widget says "out of specification but no reset", that is not a pass — it is a defect that has not been triggered yet, and Lesson 6.6 will show you what it costs to find those in the field instead of at the bench.
KEY INSIGHTHow this connects
- Passive Components for Real: Resistors, Capacitors, Inductors (1.4) — the ESL and ESR that make a capacitor stop being a capacitor were introduced there as parasitics. Here they are the specification: for decoupling, you are buying low inductance and getting capacitance as a bonus.
- Logic Levels, Pull-ups, and Open-Drain (2.1) — noise margin is measured against a ground reference, and ground bounce moves that reference. A 400 mV bounce eats most of the noise margin you carefully calculated, which is why power integrity and signal integrity are the same subject seen from two ends.
- Protection: ESD, Reverse Polarity, Overcurrent, and Inrush (2.6) — next: the same capacitors that save you from a load step become a problem at power-on, when charging them is itself a current surge. And the ESD strike that this lesson's loop inductance decides the fate of.
Everything here reappears as geometry in Stage 5: placement (5.5) is largely the art of keeping decoupling loops short, routing (5.6) is largely the art of not breaking return paths, and the stackup decision (5.4) is what made a solid layer-2 plane available in the first place. It reappears again in Stage 6 as EMC (6.4), where the same loop areas that generate voltage noise inside your board radiate it outside, and a certification lab's antenna measures exactly the quantity you have been minimising all along.
GLOSSARYKey terms introduced
- Decoupling capacitor
- A local charge reservoir at a chip's supply pin, sized to cover a load step for the interval until a larger, slower source takes over. Its usefulness is set by the inductance of its loop, not only by its value.
- ESL (equivalent series inductance)
- A capacitor's own parasitic inductance, roughly 0.7 nH for an 0402 and around 1 nH for an 0805. Added to the mounting and trace inductance to give the loop's total.
- Self-resonant frequency
- 1/(2π√(LC)) for a capacitor including its loop inductance. Below it the part is capacitive and useful; above it the part is inductive and is not decoupling.
- Anti-resonance
- The impedance peak that appears between the self-resonant frequencies of two paralleled capacitors of different values, where one is inductive and the other still capacitive. The reason a decade ladder of values can be worse than several of one value.
- Return path
- The conductor the current uses to get back to its source. Above a few kilohertz it is the copper directly beneath the outgoing trace, because that is the minimum-inductance loop.
- Loop area
- The area enclosed by the outgoing and returning current. Proportional to inductance, and hence to both the voltage generated by a current step and the field radiated by it.
- Ground bounce
- The voltage that appears across the inductance of a shared ground connection when current through it changes, shifting a chip's internal ground reference relative to the board's and eating noise margin.
- Common-impedance coupling
- Interference caused by two circuits sharing a piece of return conductor, so that one circuit's current develops a voltage in the other's ground reference.
ACTIVE RECALLCheck your understanding
Answer in your own words first, then reveal.
1A board resets when its RGB LED switches on. There is one 100 nF capacitor across the MCU's supply, mounted 3 mm away over a solid plane, and no bulk capacitor. The LED step is 60 mA. Diagnose it with numbers and give the fix.›
Separate the two mechanisms. The inductive term: 3 mm over a plane plus the part's ESL and its vias is on the order of 3.5 nH, and the LED's edge is slow (200 ns), so V = 3.5 nH × 60 mA / 200 ns ≈ 1 mV. Not the problem. The charge term: with no bulk capacitor, the 100 nF must supply 60 mA until the regulator's control loop responds, call it 5 µs. ΔV = 60 mA × 5 µs / 100 nF = 3 V. The rail collapses. The capacitor is placed perfectly and is a hundred times too small for the job it has been given.
The fix is a bulk capacitor — 10 µF near the regulator or the load — which reduces the interval the 100 nF must cover from 5 µs to a fraction of a microsecond and supplies the rest itself with a sag of 60 mA × 5 µs / 10 µF = 30 mV. Note that the diagnosis, not the reflex, chose the part: adding a second 100 nF would have changed nothing, and moving the existing one closer would have changed nothing either.
2Why does a 10 µF capacitor not simply replace both tiers? Answer in terms of impedance, with a number.›
Because above its self-resonant frequency a capacitor is an inductor. A 10 µF 0805 with about 1.2 nH of total loop inductance resonates at f = 1/(2π√(1.2 nH × 10 µF)) ≈ 1.45 MHz. At 100 MHz its impedance is 2π × 108 × 1.2 nH ≈ 0.75 Ω and rising — it is doing nothing about anything fast. A 100 nF 0402 in a 3 nH loop resonates near 9 MHz and presents about 1.9 Ω at the same 100 MHz; on its own it holds too little charge for a long event. The two are not redundant, they occupy different frequency bands, and the phrase "100 nF plus bulk" is a two-point approximation to a flat impedance curve. Note also that the numbers say the smaller capacitor's advantage at high frequency comes almost entirely from its lower ESL and smaller package, not from its capacitance — which is exactly why a badly placed 100 nF loses that advantage completely.
3A fast signal on the top layer crosses a 3 mm slot in the ground plane, cut for a row of connector pins. What physically happens, and what are the two acceptable fixes?›
Above a few kilohertz the return current wants to flow in the plane directly beneath the trace. At the slot it cannot, so it detours to whichever end of the slot is nearer and comes back — adding perhaps 20 mm of path and enlarging the loop area by orders of magnitude. Three consequences follow at once: the loop inductance rises, so V = L·di/dt generates more noise on the signal's own reference; the larger loop couples to anything else whose return shares that detour; and the loop is now an efficient radiator at the signal's harmonics, which is what an EMC pre-scan sees.
Fix one: reroute the trace so it goes around the slot rather than across it, keeping continuous copper beneath it the whole way. Fix two: remove the slot — usually possible, because slots are rarely deliberate; a row of pads becomes a slot only when the plane is not stitched between them. What is not a fix is adding a capacitor across the split: it provides a return path only at frequencies where its own impedance is low, and it introduces a resonance of its own. Treat that as a last resort on a board you cannot re-spin, not a design technique.
4Your colleague proposes splitting Beacon's ground into "analog ground" for the battery divider and "digital ground" for everything else, joined at one point near the regulator. Evaluate the proposal.›
Decline it, and be specific about why. The split's stated goal — keeping digital return current out of the analog reference — is real, but on this board it is achieved better by placement than by cutting copper. Beacon's analog node is a 500 kΩ divider filtered to 3.2 Hz by 100 nF, so it is essentially immune to the megahertz return currents the split is meant to exclude; the noise that could actually hurt it is coupled by proximity and by the ADC's own reference, neither of which a ground split addresses.
Meanwhile the split creates a concrete new hazard: any trace crossing the boundary — and on a 35 × 25 mm board with a module, two sensors, USB and an LED, something will — loses its return path and acquires the detour described in question 3. On an RF design that is the worst possible failure, because the module's antenna performance depends on a continuous reference plane (Lesson 5.7). The better proposal is one unbroken plane, the divider and its capacitor placed away from the switching regulator and the LED, and the divider's own ground via returned close to the ADC's reference. That gets the benefit the split was reaching for, with none of its risk.