EMC/EMI
EMC/EMI: The Board That Fails in the Chamber, Not on the Bench
Every board is a radio station. The only question is whether it knows it is one. A 60 mm clock trace on a 2-layer board fails radiated emissions at its 5th harmonic by 1.7 dB; the same trace on a 4-layer board passes with 14.4 dB of margin. This piece shows the three ways a board radiates, the return path that decides most failures, and the stackup that fixes them.
Every board is a radio station. The only question is whether it knows it is one.
A radio station is built to broadcast. A circuit board is built to compute, to sense, to switch, and it broadcasts anyway, because every current that flows in a loop makes a magnetic field, and every loop is an antenna. The board does not have to want to transmit. It transmits because physics does not ask permission.
On the bench, nobody notices. The board works. The sensor reads. The radio connects. The firmware runs. Then the board goes into the chamber, a room lined with absorber, with an antenna three metres away, measuring what the board is shouting into the air, and the antenna hears the clock, and the third harmonic, and the fifth harmonic, and the cable, and the slot in the ground plane. And the launch date becomes the retest date.
Example Labs is a made-up company we use for worked examples.
This is the article about why that happens, and what to change so it does not. The board is the Example Labs field sensor hub, as always, 100 × 60 × 40 mm, a CR2032 battery (230 mAh, 2,484 J) from Article #6, the enclosure from Article #7, the BLE radio from Article #12, and the demo puts it in the chamber, with the numbers worked out before writing.
What EMC is, in plain words
EMC, electromagnetic compatibility, is the design goal: the board must operate correctly without interfering with other systems, and without being affected by external noise. Two directions, one word. The board must be quiet enough not to disturb its neighbours, and thick-skinned enough not to be disturbed by them.
EMI, electromagnetic interference, is the noise itself: the unwanted electromagnetic energy that fast-switching currents, clock harmonics, or external RF sources generate. EMI is the problem. EMC is the state of not having the problem.
The board is tested in two directions, and each direction has two paths:
- Emissions, what the board gives off. Radiated emissions travel through space as electromagnetic waves. Conducted emissions travel through physical conductive paths, power supply lines, signal traces, ground loops. The chamber measures radiated emissions; a LISN (a line impedance stabilisation network, a box that holds the power line at a known impedance so the noise on it can be measured) measures conducted emissions.
- Immunity (or susceptibility), what the board can withstand. Radiated immunity is tested to IEC 61000-4-3: the board sits in a field and has to keep working. ESD, electrostatic discharge, the spark from a fingertip, is tested to IEC 61000-4-2.
The limits the board has to meet depend on where it will be sold. In the United States, FCC Part 15 Subpart B governs unintentional radiators, devices that radiate only because they are digital, not because they are transmitters. Class B is the stricter set of limits, for devices marketed for use in a residential environment, which is to say, most products.
In the European Union, the EMC Directive 2014/30/EU lists CISPR 32 as the harmonised emissions standard for multimedia equipment, covering 9 kHz to 400 GHz. The two sets of limits are close enough that a board designed to one usually meets the other, and the table in the standards section below gives both.
The important thing to understand before the rest of the piece: the bench and the chamber measure different things. The bench measures whether the board works. The chamber measures whether the board is quiet. A board can work perfectly and be loud. Most of them are.
The three ways a board radiates
A board radiates through three mechanisms, and they are not equal. One of them causes most failures.
First: the differential-mode loop. Current flows out along a trace and back along its return path. The two currents are equal and opposite, so their fields largely cancel, and what remains behaves like a small loop antenna. The field from a small loop, in free space, at distance r, is:
E ≈ 1.316 × 10⁻¹⁴ · f² · A · I / r (V/m)
where f is the frequency in hertz, A is the loop area in square metres, I is the current in amps, and r is the distance in metres. Or, in the form the EMC textbooks print, with the field in microvolts per metre, the area in square centimetres, and the frequency in megahertz:
E = 1.32 · A · I · F² / r
The f² term is the whole story: differential-mode emission rises 40 dB per decade of frequency for a constant current. Double the frequency and the field quadruples. A loop that is silent at 48 MHz is loud at 240 MHz and deafening at 528 MHz, and the demo below shows exactly that, on the same trace, with the same current.
Second: the common-mode cable. Identical currents on both conductors of a cable, or across a PCB-to-chassis boundary, create an electric monopole, and the cable is the antenna. Even microamps of common-mode current produce significant radiation, because the cable is electrically long. In practical products, common-mode radiation accounts for 80–90% of emission failures. A 5 µA common-mode current at 300 MHz on a 1-metre cable exceeds FCC Class B limits. Ten microamps on a resonant cable is generally more than enough to cause an emissions failure. The cable does not have to carry a signal. It only has to carry the noise.
Third: the slot antenna. Gaps in ground planes, connector cutouts, and board edges create slot radiators that are efficient at frequencies where the slot length is about half a wavelength. A slot is a hole that radiates, and the longer it is, the lower the frequency at which it radiates. The shield-can section below works through the numbers.
Three mechanisms, one rule that covers all three: the loop area decides. The loop area between the forward and return current, the loop area of the cable, the length of the slot. Shrink the loop and the radiation falls. Widen it and the radiation rises, with the square of the frequency.
The return path: the number-one cause
Every current flows in a loop. It goes out, and it comes back. The return path is the way back, and at high frequency it is not the way you think.
At low frequency, return current follows the path of least resistance, which might be a long, winding path through ground traces. At high frequency, return current follows the path of least inductance, which is directly beneath the signal trace, on the nearest reference plane. Above about 50 kHz, the return current prefers to flow close to the forward current, directly under the trace. The rule is the physics: inductance is minimised when the outgoing path and the returning path are as close together as possible.
So the return path is not a wire. It is a sheet, the reference plane, and the current flows in the sheet, under the trace, in a narrow band whose width is roughly the height of the trace above the plane. And that is why the first rule of EMC design is the one the Academy of EMC puts first in its guidelines:
Never route signals over split reference planes.
A split in the reference plane, a gap cut for a power island, a routing channel, a slot for isolation, forces the return current to detour around it. The loop area expands. Radiated EMI is directly proportional to that loop area.
Cuts in a reference plane can create slot antennas that actively enhance radiation. The guides are blunt about the consequence: split ground planes are the most common reason boards fail in the anechoic chamber. A gap in the return path is the number-one cause of EMI failures, and the Flux.ai guide says the same thing in the same words: most EMI failures stem from impedance discontinuities, return-path interruptions, and high-frequency edge rates that convert traces and plane gaps into unintended antennas.
The numbers make it concrete. Even a 0.5 mm routing channel across a ground plane can increase emissions by 20 dB at harmonics. A 0.5 mm cut, thinner than a pencil line, costs 20 dB, which is ten times the field. And the fix is not exotic. It is the rule, stated three ways by three guides, because it is the rule that matters:
- Keep at least one solid reference plane closely adjacent to high-frequency signals (above 1 MHz).
- Route high-speed nets directly over a continuous ground layer. Never cross a plane split.
- If a trace must change layers, the return current must change reference planes too, put a stitching via (if both planes are ground) or a stitching capacitor (if the new plane is a power plane) right next to the signal via.
If a signal absolutely must cross a split, bridge the split with stitching capacitors, 0.001 or 0.01 µF, one on either side of the trace, within 0.1 inches of the crossing, to provide an AC return path for the high-frequency content. But know what that costs: the return current must now flow through a via, a trace, a mounting pad, a capacitor, a mounting pad, a trace, and a via, 5 to 10 nH of additional inductance at minimum in the return path.
It is a repair, not a design. The design is to not split the plane.
And the plane itself: use a single continuous ground plane and physically partition the analog and digital components into separate areas of the board, rather than splitting the ground plane into separate analog and digital sections. Separate the power planes, not the ground.
The stackup: the 15 dB decision
The stackup is the order of the copper layers in the board, and it is the single cheapest EMC decision in the whole design, because it is made before the board is routed, and it decides the loop area of every trace on the board at once.
On a 2-layer board, the return current follows the path of least impedance through whatever ground copper is available, a pour, a trace, a via to the other side. Routing necessarily cuts through the ground pours, because the traces and the ground share the same two layers. Every signal trace routed across a ground pour on the same layer creates a gap that forces the return current to detour. The loop area is the trace length times the board thickness, 1.6 mm on a standard board.
On a 4-layer board with a solid inner ground plane, the return current flows directly underneath the signal trace on the adjacent plane, with no interruptions. The loop area is the trace length times the dielectric thickness between the signal layer and the plane, typically 10–12 mil, 0.25–0.30 mm. Six times smaller than the 2-layer loop. And the field follows the loop area.
The measured result, from the guides that have measured it: a four-layer board will produce 15 dB less radiation than a two-layer board, all other factors being equal. That is the rule of thumb the stackup guides print, and the measurements agree, AtlasPCB's pre-compliance measurements consistently show 10–15 dB higher radiated emissions from 2-layer boards compared to equivalent 4-layer designs operating at the same clock frequencies. A properly designed 4-layer stackup can show approximately a 10× reduction in noise compared to the same design on two layers.
The boundary is worth stating, because it is a decision rule, not a slogan. Boards operating below 25 MHz with relaxed EMC requirements, industrial environments, no FCC/CE testing, route successfully on 2 layers with careful ground-pour management. Boards that must pass FCC Class B, CE, or any radiated emissions testing with signals above 50 MHz almost always need a 4-layer stackup with unbroken ground planes. Above approximately 10 to 15 MHz, multi-layer boards should normally be considered.
And the spacing matters inside the 4-layer board too. The simplest way to improve the EMC performance of a four-layer board is to space the signal layers as close to the planes as possible (under 0.010 inches) and use a large core (over 0.040 inches) between the power and ground planes. For a 0.005-inch spacing between the trace layer and the plane layer, the reduction in trace-loop radiation can amount to 10 dB or more compared with a stackup with equally spaced layers.
The cost argument, stated plainly: a 2-layer board that should have been 4 layers costs far more in EMC re-testing, assembly rework, and schedule delays than the 30–50% premium four layers would have cost from the start. The board is cheaper. The product is not.
The local fixes: decoupling, beads, and cans
The return path and the stackup decide most of the radiation. Three local fixes handle the rest, and each one has a rule, and a failure mode when the rule is ignored.
Decoupling capacitors. A decoupling capacitor is a small capacitor placed directly across an IC's power and ground pins. It acts as a local energy reservoir, supplying instantaneous current during switching events and filtering high-frequency noise from the power rail, so the current spike stays local instead of travelling across the board and radiating from the power distribution network. The standard recommendation, in nearly every IC datasheet, is one 100 nF ceramic capacitor per IC power pin, placed as physically close as possible to the VCC and GND pins, within a few millimetres, with a target of under 2 mm from the pin and the return via within 1 mm of the capacitor's ground pad.
The reason is not the capacitance. It is the inductance. A 100 nF capacitor at 100 MHz has a capacitive reactance of 0.016 Ω, essentially a short circuit.
But the wire between the capacitor and the pin has inductance, and inductance rises with frequency. A loop of roughly 1 nH per millimetre of length gives about 2 nH at 2 mm (1.26 Ω at 100 MHz) and about 10 nH at 10 mm (6.28 Ω). The demo works the numbers out.
The point is the ratio: at 10 mm away, the inductance is 395 times the capacitance. The capacitor is a 0.016 Ω part in series with a 6.28 Ω wire, and the wire decides. A capacitor six inches from the IC is nearly useless for high-frequency decoupling, because the connecting wire adds inductance that limits the speed at which the capacitor can respond.
And above its self-resonant frequency, the capacitor's own parasitic inductance dominates and its impedance rises again, so the capacitor stops decoupling effectively. Smaller packages have lower inductance and a higher self-resonant frequency, which is why 0402 is the default for 100 nF decoupling caps. Pair the local 100 nF with a bulk capacitor (1–10 µF, or 10–470 µF electrolytic) near the power supply for the low-frequency energy, and the network covers the spectrum the single capacitor cannot.
Ferrite beads. A ferrite bead is a passive component placed in series with a power rail. At low frequencies it adds almost no resistance, current flows through with minimal loss, and a good bead has a DC resistance down to 2.5 mΩ. At higher frequencies its impedance rises sharply, typical parts are specified at 30 to 1,000 Ω at 100 MHz, and, unlike a pure inductor, the impedance at high frequency is mostly resistive. That is the difference that matters: an inductor reflects high-frequency noise back into the circuit, which can cause resonance and ringing. A ferrite bead absorbs it, converting the noise into a tiny amount of heat inside the ferrite material. It is a frequency-dependent sponge, not a wall. Paired with decoupling capacitors on the load side, the bead and the capacitors form a simple LC filter that rejects differential-mode noise on the power line. Two warnings from the guides: a high DC load current can saturate the ferrite core and remove up to 90% of its 100 MHz impedance, so the bead must be rated for the current it actually carries, and selecting the wrong part, or placing it next to an undamped capacitor, can destabilise a switching power supply.
Shield cans. A shield can is a metal can soldered over a noisy circuit, a radio, a switching regulator, a clock source, creating a localised Faraday cage: a conductive enclosure that reflects and absorbs RF, keeping internal fields in and external fields out. Above a few hundred megahertz, the metal itself is not the limit. A 0.15 mm copper-nickel-zinc wall gives about 153 dB of absorption at 1 GHz and about 48 dB at 100 MHz. Energy gets in and out through the openings, seams between wall and lid, gaps in the solder joint, ventilation holes, and every trace that crosses under the wall. The opening, not the metal, is the weak point.
The rule for openings is the wavelength. A slot behaves as a slot antenna, and the shielding effectiveness of a single slot of longest dimension L is approximately:
SE ≈ 20·log₁₀(λ / 2L) dB
When the slot length equals half a wavelength, the shield radiates and becomes useless, SE drops to zero. The design rule keeps apertures well under λ/20 at the highest frequency of concern, which gives about 20 dB of shielding effectiveness; under λ/50 is better still. And the can itself can resonate: a metal can around a noisy source traps energy inside, and if the cavity dimensions resonate at a harmonic of the noise, the enclosure becomes an antenna and radiates more efficiently than the bare board did.
A floating or single-point-grounded shield can resonate and re-radiate, making it worse than no shield at all, so the can is grounded at many points along its perimeter, with vias at λ/20 or tighter, and the lid is closed with spring fingers or a solder fillet.
The correction to the pipeline: The pipeline's note said "a slot approaching one-quarter of the target wavelength leaks". The checked sources say the slot radiates at half a wavelength (SE = 0 dB at L = λ/2), and the design rule is to keep slots under λ/20. The article uses the checked figures, and this note records the correction.
Where to look first
When a board fails radiated emissions, the frequency tells you where the radiator is.
Below 1 GHz, the radiator is usually a cable, or a unit in the system that is not grounded properly. Most radiated-emissions issues occur below 1 GHz, and the cause is usually cables or units not grounded properly. The common-mode current on the cable is the mechanism, and the fix is at the connector: a common-mode choke on every cable interface, a 360-degree bond of the cable shield to the enclosure, and a ground that is actually a ground, not a wire that happens to touch the chassis somewhere.
Above 1 GHz, the radiator is usually on the PCB. Clock lines, single-ended high-speed data lines, small PCB structures. The cable is too short, electrically, to radiate efficiently up there; the trace is not. The fix is on the board: the return path, the stackup, the clock containment, route clock traces on inner layers (stripline, enclosed between two ground planes, no radiation), keep them at least 200 mil from the board edge, and consider spread-spectrum clocking, which modulates the clock by 0.5–1.5% and reduces clock harmonics by 6–12 dB.
And the sentence that reframes the whole problem: EMC compliance is primarily a common-mode noise problem, not a signal-integrity problem. Your signals can have perfect eye diagrams while still causing emission failures. The eye diagram says the receiver understands the signal. The chamber says the rest of the world hears it too.
The standards and the test
The limits, for a Class B digital device at 3 metres, from FCC Part 15 §15.109 (checked today, from the eCFR text itself):
| Frequency | Field strength | In dBµV/m |
|---|---|---|
| 30–88 MHz | 100 µV/m | 40 |
| 88–216 MHz | 150 µV/m | 43.5 |
| 216–960 MHz | 200 µV/m | 46 |
| Above 960 MHz | 500 µV/m | 54 |
CISPR 32 Class B, at 3 metres: 30–230 MHz at 40 dBµV/m, and 230–1,000 MHz at 47 dBµV/m. The measurement is quasi-peak, a detector that weights the emission by how often it repeats, because a rare spike and a continuous tone are not the same problem, in a semi-anechoic chamber, with calibrated antennas, the unit rotated, and the antenna raised and lowered, because the maximum is what counts.
The immunity side: IEC 61000-4-3 for radiated immunity (the board in a field, still working), IEC 61000-4-2 for ESD (the spark from a fingertip, or a charged cable, and the board must not reset, corrupt, or die).
The cost of the test, from Article #15: simple DC/battery electronics, €400–€800 for EMC testing only. That is the test. The retest is the same again, plus the respin, plus the weeks. The board that fails in the chamber does not fail by €400. It fails by the respin, the retest, and the launch date that is now the retest date.
You cannot patch poor signal integrity with a metal shield at the end of the project. The shield is the last resort, not the design. By the time the board is in the chamber, the loop areas are fixed, the stackup is fixed, and the return paths are fixed, and the shield can only hide what the board is already doing. The guides say it plainly, and the article repeats it because it is the sentence that saves the most money: design for CISPR 32 and FCC Part 15 compliance from the start, because it prevents costly board respins and failed chamber tests.
The hub in the chamber
The Example Labs field sensor hub, as always: 100 × 60 × 40 mm, a CR2032 battery (230 mAh, 2,484 J) from Article #6, the enclosure from Article #7, the BLE radio from Article #12. The hub's clock runs at 48 MHz. The hub has a 60 mm clock trace, a 1-metre sensor cable, and a shield can over the radio with a 30 mm ventilation slot. The demo puts the hub in the chamber, three metres from the antenna, and works out what the antenna hears.
Every figure below is the article's own arithmetic, run in Python before writing, from the loop-antenna formula the EMC textbooks print (E = 1.32 · A · I · F² / r, with E in µV/m, A in cm², I in amps, F in MHz, r in metres) and the slot formula (SE ≈ 20·log₁₀(λ / 2L)).
Step 1: The clock trace: the loop that radiates
The 60 mm clock trace, carrying 10 mA of harmonic current, at the 5th harmonic of the 48 MHz clock, 240 MHz.
On the 2-layer board, the trace sits 1.6 mm above the ground pour (the board thickness), so the loop area is 60 mm × 1.6 mm = 0.96 cm². On the 4-layer board, the trace sits 0.25 mm above the ground plane (10 mil of prepreg), so the loop area is 60 mm × 0.25 mm = 0.15 cm². The ratio is 6.4×, 16.1 dB, which is the loop-area model's version of the 15 dB rule of thumb, and the two agree within a decibel.
| Frequency | 2-layer field | Limit | 2-layer verdict | 4-layer field | 4-layer margin |
|---|---|---|---|---|---|
| 48 MHz (fundamental) | 9.7 µV/m (19.8 dBµV/m) | 100 µV/m (40) | passes by 20.2 dB | 1.5 µV/m (3.6) | 36.4 dB |
| 144 MHz (3rd harmonic) | 87.6 µV/m (38.8) | 150 µV/m (43.5) | passes by 4.7 dB | 13.7 µV/m (22.7) | 20.8 dB |
| 240 MHz (5th harmonic) | 243.3 µV/m (47.7) | 200 µV/m (46) | fails by 1.7 dB | 38.0 µV/m (31.6) | 14.4 dB |
| 336 MHz (7th harmonic) | 476.9 µV/m (53.6) | 200 µV/m (46) | fails by 7.5 dB | 74.5 µV/m (37.4) | 8.6 dB |
| 528 MHz (11th harmonic) | 1,177.6 µV/m (61.4) | 200 µV/m (46) | fails by 15.4 dB | 184.0 µV/m (45.3) | 0.7 dB |
The fundamental passes easily. The 3rd harmonic passes by less than 5 dB. The 5th harmonic fails, by 1.7 dB, which is how most boards fail: not by 20 dB, but by a decibel or two, on a harmonic nobody was watching. The 7th fails by 7.5 dB, and the 11th by 15.4 dB. The same trace, the same current, the same board, and the field grows with the square of the frequency, exactly as the formula says.
On the 4-layer board, the 5th harmonic passes with 14.4 dB of margin, and the 7th with 8.6 dB. But note the 11th harmonic: even the 4-layer board passes by only 0.7 dB. The stackup alone does not fix everything. The current spectrum matters too, the edge rate, the harmonic content, and that is where spread-spectrum clocking (6–12 dB off the harmonics) and a slower edge rate earn their keep.
Two honest caveats, stated in the demo rather than hidden: the loop formula is a free-space model, and over a reflecting ground plane, as on an open-area test site, the reflected wave can add up to 6 dB, which would turn the 5th-harmonic fail from 1.7 dB into 7.7 dB. And the harmonic current is assumed at 10 mA for every harmonic, which is a simplification, real harmonic currents fall with harmonic number for a trapezoidal waveform, so the 11th harmonic's current would be lower than the 5th's.
The point that does not depend on either caveat is the 16.1 dB between the two stackups, which is a ratio of loop areas, not of currents.
Step 2: The sensor cable: the 5 microamps that fail
The hub's 1-metre sensor cable, carrying 5 µA of common-mode current at 300 MHz. The common-mode field from a short cable is:
E ≈ 1.257 × 10⁻⁶ · f · L · I / r (V/m)
At 300 MHz, 1 metre, 5 µA, 3 metres: 628.5 µV/m (56.0 dBµV/m), against a limit of 200 µV/m (46 dBµV/m). It fails by 10 dB. Five microamps. Not five milliamps, five microamps, a current so small that no multimeter on the bench would ever see it, and the chamber hears it at three metres.
The fix is not on the board. It is at the connector: a common-mode choke on the cable interface, and a 360-degree bond of the cable shield to the enclosure. This is the below-1-GHz radiator, and it is the one the board designer forgets, because the cable is not on the board.
Step 3
The hub's shield can over the BLE radio has a 30 mm ventilation slot. The BLE radio is at 2.4 GHz, where the wavelength is 125 mm. The slot's shielding effectiveness is:
SE ≈ 20·log₁₀(λ / 2L)
- A 30 mm slot at 2.4 GHz: SE = 20·log₁₀(125 / 60) = 6.4 dB, nearly useless. The can is a metal box with a hole in it, and the hole is doing the talking.
- Five 5 mm slots instead of one 30 mm slot: SE = 20·log₁₀(125 / 10) = 21.9 dB, the same total ventilation, 15.5 dB better, because the longest dimension decides, not the area.
- The same 5 mm slot at 5.8 GHz (the Wi-Fi band): SE = 14.3 dB, because the wavelength is shorter (51.7 mm) and the slot is relatively longer.
- A 62.5 mm slot at 2.4 GHz, exactly half a wavelength: SE = 0 dB. The shield radiates and becomes useless.
The design rule, from the checked sources: keep every aperture under λ/20 at the highest frequency of concern, 6.25 mm at 2.4 GHz, which gives about 20 dB of shielding effectiveness. The fix is not a thicker can. The metal already gives 153 dB at 1 GHz. The fix is smaller holes, more of them, and a lid that is actually closed.
Step 4: The decoupling capacitor that is too far away
The hub's MCU has a 100 nF decoupling capacitor on each power pin. On the first board, the capacitors sit 10 mm from the pins, because the routing was easier that way. On the second board, they sit 2 mm away.
A 100 nF capacitor at 100 MHz has a capacitive reactance of 0.016 Ω. But the loop between the capacitor and the pin has inductance, roughly 1 nH per millimetre of loop length, and inductance rises with frequency:
| Distance | Loop inductance | Inductive reactance at 100 MHz | Inductance ÷ capacitance |
|---|---|---|---|
| 2 mm | ~2 nH | 1.26 Ω | 79× |
| 5 mm | ~5 nH | 3.14 Ω | 197× |
| 10 mm | ~10 nH | 6.28 Ω | 395× |
| 25 mm (1 inch) | ~25 nH | 15.71 Ω | 987× |
At 10 mm away, the capacitor is a 0.016 Ω part in series with a 6.28 Ω wire, and the wire decides. The capacitor is not decoupling anything at 100 MHz, the inductance is. Move it to 2 mm and the impedance falls fivefold, to 1.26 Ω. The capacitor did not change. The distance did. That is the whole lesson of decoupling: place the capacitor, and the placement is the component.
The cost of the failure
The hub failed at 240 MHz by 1.7 dB, and at 336 MHz by 7.5 dB. The test cost €400–€800 for EMC testing only (from Article #15, for simple DC/battery electronics). The retest costs the same again, plus the respin, and the respin is the 4-layer board the hub should have been from the start, at a 30–50% premium on the bare board, against the retest, the rework, and the launch date that is now the retest date. The 4-layer board was the cheap option. It always was.
Decision test: four questions about your board
- Is there a solid reference plane directly under every high-speed trace, and does any trace cross a split in it? If the answer to the second half is yes, the board has a slot antenna, and the chamber will find it. The plane is the return path, and the return path is the radiation.
- Is the board 4 layers, with the signal layer within 10 mil of the plane, and if it is 2 layers, is every signal below 25 MHz? If the board has signals above 50 MHz and has to pass FCC Class B or CE, and it is 2 layers, the chamber will decide the stackup for you, at the retest.
- Is there a 100 nF capacitor within 2 mm of every IC power pin, and a ferrite bead on every noisy power rail, rated for the current it actually carries? If the capacitors are 10 mm away, they are not decoupling. If the bead saturates at the load current, it has lost up to 90% of its impedance, and the rail is unfiltered.
- Is every aperture in the shield under λ/20 at the highest frequency of concern, and is every cable interface choked and bonded? If the slot is 30 mm at 2.4 GHz, the can is a box with a hole in it. If the cable has 5 µA of common-mode current at 300 MHz, the cable is the antenna, and the board is not the problem.
Objections
"But it works on the bench." The bench measures whether the board works. The chamber measures whether the board is quiet. A board can work perfectly and be loud, and most of them are. The bench has no antenna three metres away, and no limit. The chamber has both.
"But we'll add a shield can at the end." You cannot patch poor signal integrity with a metal shield at the end of the project. The shield is the last resort, not the design. By the time the board is in the chamber, the loop areas are fixed, the stackup is fixed, and the return paths are fixed, and the shield can only hide what the board is already doing, through holes that are themselves antennas. Design for CISPR 32 and FCC Part 15 from the start.
"But 4 layers cost more." The 4-layer board costs 30–50% more than the 2-layer board. The 2-layer board that should have been 4 layers costs the retest, the respin, the rework, and the launch date. The board is cheaper. The product is not.
"But our clock is only 48 MHz." The clock is 48 MHz. The 5th harmonic is 240 MHz, and it fails by 1.7 dB. The 11th harmonic is 528 MHz, and it fails by 15.4 dB. The field grows with the square of the frequency, so the harmonic that matters is never the fundamental. It is the fifth, the seventh, and the eleventh, the ones nobody was watching.
Sources
- Academy of EMC, "EMC Design Guidelines" (academyofemc.com)
What we used
Guideline #1: never route signals over split reference planes; the return current above ~50 kHz prefers to flow directly under the forward current; the rule of thumb: at least one solid reference plane closely adjacent to high-frequency signals (f > 1 MHz); Guideline #2: always consider the return current and minimise the loop area; "Most Common EMC Issues": most radiated-emissions issues occur below 1 GHz, where the causes are usually cables or units not grounded properly, and above 1 GHz the radiator is most probably on the PCB (clock lines, single-ended high-speed data lines, small PCB structures); radiated emissions testing 30 MHz to 6 GHz per CISPR 32 or CISPR 11; radiated immunity per IEC 61000-4-3; Guideline #7: keep clock lines as short as possible; Guideline #16: minimise the loop area of signals in cables
- Flux.ai, "EMI/EMC PCB Design Guide: How to Reduce Interference (2026)"
What we used
"Every gap in a return plane is a slot waiting to radiate"; a continuous ground plane and tightly controlled return paths are the single most effective way to reduce EMI before shielding or filtering; CISPR 32 covers 9 kHz to 400 GHz; current must always return to its source, and the return path defines the loop area, and loop area defines how much the board radiates; route adjacent to reference planes, high-speed nets must run directly over a continuous ground layer, never cross a plane split; "Split ground planes are the most common reason boards fail in the anechoic chamber"; cuts in a reference plane can create slot antennas that actively enhance radiation; if a split must be bridged, use stitching capacitors across the gap; the 3W rule for trace spacing
- Flux.ai, "PCB EMI Issues: Causes, Common Mistakes, and Layout Solutions (2026)"
What we used
"Most EMI failures stem from impedance discontinuities, return path interruptions, and high-frequency edge rates that convert PCB traces and plane gaps into unintended antennas"; "A gap in the return path forces current to find an alternative route, creating a large loop antenna. This is the number one cause of EMI failures."
- AtlasPCB, "2 Layer vs 4 Layer PCB: Engineering Decision Guide"
What we used
4-layer signal-to-plane distance 10–12 mil (0.25–0.30 mm); "we consistently see 10–15 dB higher radiated emissions from 2-layer boards compared to equivalent 4-layer designs operating at the same clock frequencies"; boards below 25 MHz with relaxed requirements route successfully on 2 layers; boards that must pass FCC Class B, CE, or any radiated emissions testing with signals above 50 MHz almost always need a 4-layer stackup; a 2-layer board that should have been 4 layers costs far more in EMC re-testing, assembly rework, and schedule delays than the 30–50% premium four layers would have cost from the start
- SMTnet / Gold Phoenix PCB, "PCB Stack-Up"
What we used
"A rule of thumb, that is often used, is that a four-layer board will produce 15 dB less radiation than a two-layer board, all other factors being equal"; for 0.005-inch spacing between the trace layer and the plane layer, the reduction in trace-loop radiation can amount to 10 dB or more compared with a stackup with equally spaced layers; above approximately 10 to 15 MHz, multi-layer boards should normally be considered; the simplest way to improve a four-layer board is to space the signal layers as close to the planes as possible (<0.010 inches) and use a large core (>0.040 inches) between power and ground
- AtlasPCB, "EMC Shielding Techniques in PCB Layout"
What we used
"Even a 0.5 mm routing channel across a ground plane can increase emissions by 20 dB at harmonics"; use a unified ground plane and partition analog and digital components into separate areas rather than splitting the ground plane; the 20H rule for plane edges
- AtlasPCB, "PCB Design for EMC Pre-Compliance"
What we used
Differential-mode radiation is proportional to loop area × frequency² × current; common-mode radiation accounts for 80–90% of emission failures; a 5 µA common-mode current at 300 MHz on a 1-metre cable exceeds FCC Class B limits; clock harmonics are the strongest source of narrowband emissions; spread-spectrum clocking (0.5–1.5% modulation) reduces clock harmonics by 6–12 dB; decoupling capacitor target: under 2 mm from the IC power pin, with the return via within 1 mm of the capacitor's ground pad; route clock traces on inner layers (stripline) whenever possible, and keep them at least 200 mil from the board edge; common-mode chokes on every cable interface; ground-plane stitching vias along the board perimeter at λ/20 spacing
- Hott / Shield Digital Design, "PCB Stack-Up, Part 6: Return Path Discontinuities"
What we used
The lowest-impedance return path is in a plane directly underneath the signal trace; stitching capacitors across a split should be 0.001 or 0.01 µF, within 0.1 inches of the crossing, and add 5 to 10 nH minimum inductance in the return path; interplane capacitance is too small to be effective except above ~500 MHz; if both power and ground planes are split, no signal traces should cross the split boundary
- eCFR, 47 CFR Part 15 §15.109 (checked today, primary source)
What we used
Radiated emission limits for Class B digital devices at 3 metres: 30–88 MHz at 100 µV/m, 88–216 MHz at 150 µV/m, 216–960 MHz at 200 µV/m, above 960 MHz at 500 µV/m; Class B is for devices marketed for use in a residential environment
- TPS Elektronik, "How to Pass CISPR 32 and FCC Part 15 Radiated Emissions Testing"
What we used
CISPR 32 Class B at 3 metres: 30–230 MHz at 40 dBµV/m, 230–1,000 MHz at 47 dBµV/m; FCC Part 15 Class B at 3 metres: 40 dBµV/m (30–88 MHz), 43.5 dBµV/m (88–216 MHz), 46 dBµV/m (216–960 MHz), 54 dBµV/m (above 960 MHz); the EMC Directive 2014/30/EU lists CISPR 32 as the harmonised emissions standard for multimedia equipment
- RFFngineer, "EMC and EMI for RF Designers"
What we used
The small-loop formula E ≈ 1.316 × 10⁻¹⁴ · f² · A · I / r (V/m), with the f² term meaning differential-mode emission rises 40 dB per decade; over a reflecting ground plane the reflected wave can add up to 6 dB; the worked example: a 20 mA harmonic at 100 MHz in a loop of 2 cm² gives 175 µV/m, or 44.9 dBµV/m; the common-mode cable formula E ≈ 1.257 × 10⁻⁶ · f · L · I / r (V/m)
- The Free Library, "Estimating emissions from your printed circuit board" (Kraus)
What we used
E = 1.32 · A · I · F² / r, with E in µV/m, A in cm², I in amps, F in MHz, r in metres; the worked example: 30 MHz, a 25 cm² loop, 48 mA gives 475 µV/m against a Class B limit of 100 µV/m; the same loop over a ground plane gives 14.3 µV/m; accurate to within 1% for loops up to 10 × 10 inches, below resonance
- LearnEMC, "EMC Question of the Week: September 1, 2025"
What we used
The common-mode current on cables produces most radiated emissions from table-top products below a few hundred megahertz; ten microamps of common-mode current on a resonant cable is generally more than enough to cause an emissions failure
- Cadence System Analysis, "The Size of Apertures and the EMI Shield Wavelength"
What we used
When the dimension of an opening in an EMI shield is equal to one-half the wavelength of interference, the shield radiates EMI and becomes useless; the ideal aperture dimension is λ/20, giving 20 dB of shielding effectiveness; keep apertures under λ/50 where possible, with λ/20 as the maximum
- RFFngineer, "RF Shielding Cans and Isolation"
What we used
The slot formula SE ≈ 20·log₁₀(λ / 2L) dB, valid while L is less than half a wavelength; a 5 mm gap gives about 21.9 dB at 2.4 GHz but only 14.3 dB at 5.8 GHz; ten such gaps along a lid seam remove another 10 dB; via spacing under the wall at λ/20 or tighter; a can whose first resonance falls in an operating band can make isolation worse than no can at all; a copper-nickel-zinc can, 0.15 mm wall, gives about 153 dB of absorption at 1 GHz and about 48 dB at 100 MHz, above a few hundred megahertz the metal is not the limit, the openings are; a floating or single-point-grounded shield can resonate and re-radiate
- Ziitek, "EMI Shielding Basics for Electronics Engineers"
What we used
A metal can around a noisy source traps energy inside, and if the cavity resonates at a harmonic of the noise, the enclosure becomes an antenna and radiates more efficiently than the bare board did; a can bonded at many points along its perimeter behaves like a wall, while one grounded at a couple of tabs behaves like a patch antenna with a deliberate resonator inside
- Lighthouse LEDs, "Decoupling Capacitors: Eliminating IC Power Supply Noise"
What we used
The standard recommendation in nearly every IC datasheet is one 100 nF ceramic capacitor per IC power pin, placed as physically close as possible; a capacitor six inches from the IC is nearly useless for high-frequency decoupling; on a PCB, decoupling capacitors should be within a few millimetres of the pin; the 100 nF ceramic handles high-frequency noise, the 1–10 µF electrolytic handles low-frequency noise
- JLCPCB, "Decoupling Capacitor Guide: Values, Placement & Layout"
What we used
Place one decoupling capacitor near each IC's power pin; use 100 nF X7R MLCCs for most digital circuits; above its self-resonant frequency the parasitic inductance dominates and impedance rises again, so the capacitor stops decoupling effectively; smaller packages have lower ESL and a higher SRF
- RayPCB, "100 nF (0.1 µF) Capacitor"
What we used
The 100 nF capacitor is the most frequently specified decoupling capacitor in electronics design; X7R is the default dielectric; 0402 is the default package for automated SMT, because the small footprint keeps the cap physically close to the IC pin, which directly reduces parasitic loop inductance
- Electronic Design / Bourns, "High-Current Chip Ferrite Beads for High-Density Power" (June 2026)
What we used
Ferrite beads filter high-frequency power-supply noise while enabling multiple power rails to share the same voltage source; at lower frequencies a bead adds almost no resistance, and at higher frequencies its impedance increases sharply; impedances from 30 to 1,000 Ω at 100 MHz; DC resistance down to 2.5 mΩ; placed in series with the power rail, paired with decoupling capacitors to form an LC filter
- EDN, "The secret life of the ferrite bead"
What we used
At low frequencies the bead behaves like an inductor; as frequency climbs into the megahertz range, the resistive component takes over, and instead of reflecting high-frequency noise back into the circuit, the bead absorbs it, dissipating it as a tiny amount of heat; the parasitic capacitor can provide a bypass path at very high frequencies, limiting effectiveness
- PCBark, "Ferrite Bead on a PCB"
What we used
A high DC load current can saturate the ferrite core and remove up to 90% of its 100 MHz impedance; selecting the wrong part or positioning it next to an undamped capacitor can destabilise switching power supplies (Analog Devices AN-1368: a 10 dB gain at the 2.4 MHz ripple frequency)
- Correction recorded in this file: the pipeline's "a slot approaching one-quarter of the target wavelength leaks"
What we used
Is corrected, the checked sources say a slot radiates at half a wavelength (SE = 0 dB at L = λ/2, per Cadence and RFFngineer), and the design rule is to keep apertures under λ/20 (about 20 dB of shielding effectiveness). The article uses the checked figures
- Demo arithmetic: the article's own, run in Python before writing (60 mm trace; 2-layer loop area 0.96 cm² at 1.6 mm, 4-layer 0.15 cm² at 0.25 mm, ratio 6.4× = 16.1 dB; loop formula E = 1.32·A·I·F²/r at 10 mA and 3 m
What we used
48 MHz: 9.7 vs 1.5 µV/m; 144 MHz: 87.6 vs 13.7; 240 MHz: 243.3 vs 38.0 (2-layer fails the 200 µV/m limit by 1.7 dB); 336 MHz: 476.9 vs 74.5; 528 MHz: 1,177.6 vs 184.0; common-mode cable at 300 MHz, 1 m, 5 µA: 628.5 µV/m vs the 200 µV/m limit; slot SE at 2.4 GHz: 30 mm → 6.4 dB, 5 mm → 21.9 dB, 62.5 mm (λ/2) → 0 dB; at 5.8 GHz, 5 mm → 14.3 dB; 100 nF Xc at 100 MHz: 0.016 Ω; loop inductance ~1 nH/mm, 2 mm → 1.26 Ω, 10 mm → 6.28 Ω, 25 mm → 15.71 Ω)
- No Senova results, products, or client numbers stated anywhere in this draft. No Senova prices in any draft.