The Rail That Doesn't Ring
The Rail That Doesn't Ring
Your schematic is correct and your board still fails EMC. The problem is in the layout: the hot loop, the switch node, and the ground that is not one ground. Here is the buck converter's noise, measured in numbers.
Picture the board on the bench. The schematic is correct, you checked it three times, and the simulation agrees. The analog readings from the sensor are noisy anyway: the same reading, taken twice, comes back different by more than the sensor's own drift.
The board fails radiated emissions in the chamber, and the digital circuits crash when the motor driver switches. The schematic is correct, and the board still does not work. The problem is not in the circuit.
It is in the layout, in the copper, in the loop the switching currents travel in, and in a ground that is one symbol on the schematic and two different voltages on the board.
This article is about the noise a switching power supply makes, and the three places it comes from: the ripple on the output, the loop that radiates it, and the ground that carries it into everything else.
Ripple: the small AC voltage on top of the DC
"Voltage ripple refers to the small AC voltage that sits on top of the regulated DC output. The goal is to minimize the amplitude of this voltage ripple, to present as clean of a DC source as possible to the rest of the circuit." (EEPower) A power supply with too much noise can crash digital circuits or add noise to the output of analog circuits.
The switching frequency is the giveaway: "If the noise matches the switching frequency of your switching converter that's already a pretty clear sign of who the guilty party is." (Power Electronics News)
How much ripple is too much depends on what the supply is powering. "The threshold of ripple depends on what the supply is powering. 50mV to 100mV is likely an acceptable range for most devices, but some are more sensitive." (EEPower) The worked example is an A7 processor core running at 900 MHz: the VDD_ARM_CAP rail should be set to 1.275 V, "however, it will work at 1.3V (25mV higher) or 1.25V (25mV lower).
The acceptable voltage ripple on this rail is 25mV." A 25 mV budget on a 1.275 V rail is a 2% window, and the ripple is not the only noise on the rail. The spikes that ride on top of the ripple, at the switching edges, are faster and taller, and the ripple budget has to cover them too.
There are two kinds of noise on the rail, and they need different fixes. Ripple is the slow one: the triangular rise and fall of the inductor current, at the switching frequency, flowing through the output capacitor. Its size is set by the inductor ripple current and the output capacitor, ΔV ≈ ΔIL × ESR + ΔIL / (8 × fSW × C), and the fix is a bigger inductor, a lower-ESR capacitor, or more capacitance.
The spikes are the fast one: the switching edges, with rise times under 10 ns, ringing against the parasitic inductance of the traces and the capacitor. Their size is set by V = L·di/dt, and the fix is layout, shorter traces, a smaller loop, a smaller switch node. Adding capacitance does not fix the spikes, because at the switching frequency the capacitor's ESR matters more than its capacitance: "For oscillations of the same frequency as the switching (400kHz), in this case the ESR of the capacitors is more important than their capacitance.
What is the series resistance of capacitors? If you multiply that resistance by the variation in the inductor current, you will probably find the ripple voltage as solution." (EEVblog)
And there is a trap in the capacitance itself. A ceramic capacitor does not hold its rated capacitance at its rated voltage. In the EEVblog thread, a 50 V 10 µF 0603 capacitor "is like 1.5uF at 24V due to DC bias."
The schematic says 22 µF; the board has 3 µF. The ripple the designer calculated is not the ripple the board has. Use X7R or X5R dielectrics, and never Y5V, which "loses 80% capacitance at rated voltage" (AndwinPCB).
And measure the ripple properly: "First sanity-check the measurement: use a tip-and-barrel probe right at the output cap, those spikes often come from long probe grounds." A long probe ground is an antenna, and the 1 MΩ impedance of the probe "will pick up EMC from the air, and I found that often to be several times the actual noise." The right tool is a power rail probe, with 50 kΩ DC and 50 Ω AC controlled impedance, or, failing that, the spring tip that came with the probe, and the probe's ground on the output capacitor's own ground pad, not on a ground three centimetres away.
The hot loop: the loop that radiates
"Proper layout is critical for optimal performance in a SMPS circuit. Rapidly switching currents generate an alternating magnetic field, which may have some adverse effects. The loop that the switching currents in an SMPS travel in is referred to as the 'hot loop.' We seek to minimize the size of this loop to limit the EMI being generated by the switching currents." (EEPower)
For a buck converter, the loop is easy to draw and easy to get wrong. "The hot loop for a buck converter circuit travels through the inductor, through the output capacitor, and back through the ground plane to the device ground and input capacitor ground. The simplest way to keep the hot loop as small as possible is to keep the three grounds (input cap, device, and output cap) as close together as possible.
They should be connected with a solid ground plane on the same layer as the components, such that the path is as short as possible, and the violent switching currents avoid the internal PCB ground plane entirely." (EEPower)
The same point, from Linear Technology's application note (Kueck, AN139): only one loop in the buck carries a fully switched AC current, "switched from zero to IPEAK and back to zero. We refer to the green loop as a hot loop, because it has the highest AC and EMI energy." The other two loops, the on-cycle loop and the off-cycle loop, carry trapezoidal currents, and their combination is a smooth triangle.
The hot loop is the difference between them. "To reduce EMI and improve functionality, you need to reduce the radiating effect of the green loop as much as possible. If we could reduce the PC-board area of the green loop to zero and buy an ideal input capacitor with zero impedance, the problem would be solved."
You cannot. You can only make the loop small, and the input capacitor good.
The hot loop is defined the same way in the parasitics literature: "The hot loop of a switching-mode power converter is defined as the critical high frequency (HF) AC current loop formed by the HF capacitor and adjacent power FETs. It is the most critical part of the power stage PCB layout because it contains high dv/dt and di/dt noisy content. A poorly designed hot loop layout suffers from a high level of PCB parasitic parameters, including the ESL, ESR, and equivalent parallel capacitance (EPC)."
(EEPower) The high di/dt through the ESL causes high-frequency ringing, and the ringing is the EMI. The energy stored in the ESL is dissipated on the ESR, which is extra power loss. Minimising the loop's ESR and ESL improves efficiency, lowers the ringing, and reduces the emissions, three wins from one fix.
The rule, stated as a number: position the input bulk capacitors within 5 mm of the switching IC's VIN and PGND pins (AndwinPCB). "Every millimeter adds inductance converting switching current into voltage spikes. For 5A switching in 8ns, even 5nH generates a 3V spike from L(di/dt)."
And the field experience behind the rule: "Power supply failures account for 35% of PCB-related field returns, most from layout mistakes. When designing buck converters or boost regulators above 100 kHz, layout becomes the circuit, not just component arrangement." The author had "debugged supplies that met electrical specs but radiated 15 dB over limits or exhibited 200 mV ripple instead of 20 mV.
The difference? Layout decisions that turned parasitic inductance into failures."
The switch node: the antenna on the board
"It is also important to keep the switch node as small as possible. In a buck converter, this is the trace between the IC and the inductor, where currents are switched on and off at the switching frequency. This node is the primary source of EMI. The switch node behaves like a radiating antenna and by limiting the physical dimensions of the trace, the generated emissions can be limited." (EEPower)
The switch node is the one place on the board where the voltage slews at 5–20 V/ns (AndwinPCB), carrying full load current, and it radiates broadband noise from tens to hundreds of megahertz. The rule, stated as a number: keep switching node copper under 20 mm² for designs under 5 A, and under 50 mm² for 10 A and above. Place the inductor within 3 mm of the SW pin on integrated-FET controllers.
Never extend the switch node trace to other board sections, and never route it near analog circuitry, the capacitive coupling injects the switching noise straight into the victim. Use ground plane cutouts beneath the switch node to prevent capacitive coupling that increases common-mode EMI, with a 2 mm clearance.
The two rules for a buck converter, from EDN, are the whole layout discipline in two lines: reduce the surface area of nodes with quickly changing voltages (high dv/dt), and reduce the area of current loops with quickly changing currents (high di/dt). The switch node is the first rule. The hot loop is the second. Everything else in the layout, the input capacitor placement, the ground plane, the feedback routing, is a consequence of those two.
Input vs. output: the asymmetry that decides where to filter
"One thing about buck converters is that they have much worse emi at the input than the output. The output current is triangular while the input is has discontinuous jumps. So pay particular attention to layout and filtering at the input." (EEVblog)
The asymmetry is in the shape of the current. The output current is the inductor current: a triangle, rising and falling smoothly, never disconnected. The input current is a square wave: it rises and falls "resembling a square wave as it is disconnected from the input each switching cycle" (Vicor), jumping from zero to the peak inductor current and back to zero, twice per cycle.
And the shape decides the harmonics: "Square wave, harm(n) α 1/n. Triangle wave, harm(n) α 1/n²." (Vicor) The square wave's harmonics fall slowly, one over n; the triangle's fall fast, one over n squared.
At the 10th harmonic, the input still carries a tenth of the switching current, while the output carries a hundredth of the ripple. At the 60th harmonic, 30 MHz at a 500 kHz switching frequency, inside the FCC's radiated band, the input carries a sixtieth, the output a thirty-six-hundredth. The input is 182 times noisier than the output at that frequency.
The same point, from TI's application note (SNVAA93): "The switcher's input current iQ1 is discontinuous resulting in fundamental and lower-order harmonics as well as high frequency noise current. Differential noise is caused by the opposite propagation paths between the positive and return power lines and is related to the fast di/dt of the input current." And the input loop can bypass the filter entirely: "the input current loop can induce magnetic field that bypasses the EMI filter and couples to the LISN... it is critical to minimize the input current loop of the buck converter by strategically placing the input capacitor to reduce magnetic coupling, and it is also beneficial to use a fully sealed inductor to reduce magnetic coupling."
The input capacitor is the part that has to absorb this. Kueck's rule: "the supply wire as well as the load wire should not have high AC current content. So, we can focus our analysis from the input capacitor, CIN, which should source all relevant AC currents to the output capacitor, COUT, where any AC currents end."
The input capacitor's RMS current is IOUT × √(D(1−D)), at a 20.8% duty cycle and 2 A of load, 0.81 A RMS, flowing through the capacitor's ESR and ESL every cycle. If the input capacitor is far from the IC, or its ground returns through the internal ground plane instead of straight to the IC's PGND pin, that 0.81 A flows through the board's ground, and the board's ground is the next section's problem.
Ground: the node that is not a node
"DC/DC converter ground noise is usually a voltage developed across real impedance, not noise generated by an abstract ground node. Fast switching currents flow through trace, via, package, capacitor, connector, and plane inductance, producing V = L·di/dt." (RottenWifi)
Two PCB locations can share the same schematic ground symbol while having different instantaneous voltages. A switching current pulse through copper creates a resistive drop; through parasitic inductance it creates a much faster spike. The resistive part is small and the inductive part is not: "A 1-ounce copper plane is approximately 500 µΩ per square, so a 1 A change can produce about 500 µV per square of resistive drop."
(RottenWifi, corroborated by EDN: "The sheet resistance of 1-oz copper is about 500 µΩ/□, so a 1A change in current produces 500 µV/□ of bounce, a problem for thin, long, or daisy-chained grounds or precision electronics.") The inductive part is the one that hurts: "Ground bounce is the amount that a ground return rises or falls relative to the system's 0V reference, and, in a dc/dc-switching converter, ground bounce can be many volts, often because of changing magnetic flux."
(EDN) "Generally, pc-board-ground-plane resistance is a less important source of ground bounce than magnetic-flux change... Therefore, the best way to reduce ground bounce in a switching dc/dc converter is to control changes in magnetic flux." Control the flux, and you control the bounce.
The flux is the loop area. The loop area is the layout.
The fix has two parts, and they are the same two parts as the hot loop. First, keep the switching currents out of the rest of the board's ground: the three grounds (input cap, device, output cap) close together on a solid ground plane on the same layer as the components, "such that the violent switching currents avoid the internal PCB ground plane entirely" (EEPower). Second, do not split the ground plane beneath the converter: "Do not split or cut the ground plane beneath the DC/DC converter circuit.
Instead, use a solid GND plane, which most effectively minimizes the loop areas of critical current paths." (Academy of EMC) A split ground plane is a slot antenna, Article #19's slot, which radiates when its length approaches half a wavelength, and a split under the converter is a slot in exactly the place the switching currents are strongest. And do not let the analog and noisy returns intermingle: "If you simply place vias into the ground plane at the negative connections for the loads of analog and noisy then you run a serious risk that the return currents will intermingle."
(Power Electronics News) The high-frequency return current follows the path of least inductance, which is the smallest loop, directly under the signal trace. Give the analog return its own path back to the converter's ground, and it will take it. Give it a via into a shared plane, and the switching current will take it instead.
Example Labs is a made-up company we use for worked examples.
Example Labs' sensor hub board, 24 V to 5 V
Example Labs' field sensor hub board is the cluster's board: 100 × 60 mm, 4-layer, 120 nets, 114 test points (from Article #32), fed from the 48 V bench pack (13S1P NMC, 54.6 V full, from Articles #20 and #23) through a buck converter that steps 24 V down to 5 V at 2 A, switching at 500 kHz, the FanyPCB worked example, re-run here in Python before writing. The board carries the BLE radio at 2.4 GHz (from Article #12) and the sensor whose readings the noise corrupts (from Article #17).
Here is the rail, measured in numbers.
Step 1, the inductor and the ripple current. The duty cycle is D = VOUT / VIN = 5 / 24 = 0.208. The inductor ripple current is ΔIL = (VIN − VOUT) × D / (L × fSW) = (24 − 5) × 0.208 / (10 µH × 500 kHz) = 0.79 A peak-to-peak, 39.6% of the 2 A load current, at the top of the 20–40% rule of thumb (TI SLVA477). The peak inductor current is IL,peak = IOUT + ΔIL / 2 = 2.40 A. The inductor has to be rated above that, with margin, and the input capacitor has to carry an RMS current of IOUT × √(D(1−D)) = 0.81 A every cycle. Swap the 10 µH inductor for 22 µH and the ripple falls to 0.36 A, the peak to 2.18 A, the same load, half the ripple, for a bigger part and a slower transient response. That is the trade, and it is the first decision in the demo.
Step 2, the output ripple, and which capacitor is on the board. The ripple voltage is ΔV = ΔIL × ESR + ΔIL / (8 × fSW × C). With a ceramic output capacitor (ESR 5 mΩ, a stated assumption) at 22 µF: 4.0 mV from the ESR plus 9.0 mV from the capacitance, 13.0 mV total, inside the 50–100 mV range, and inside the 25 mV A7 limit. With an electrolytic output capacitor (ESR 100 mΩ, a stated assumption) at the same 22 µF: 79.2 mV from the ESR plus 9.0 mV from the capacitance, 88.2 mV total, inside the 50–100 mV range, but 7 times worse, and the ESR term is 9 times the capacitance term. That is the EEVblog point, in numbers: at the switching frequency, the ESR matters more than the capacitance. And with the DC-bias trap, the ceramic capacitor that is 22 µF on the schematic and 3 µF on the board at 24 V, the capacitance term alone becomes 66.0 mV, for 69.9 mV total. The schematic's ripple calculation is not the board's ripple. The board's ripple is the one the capacitor actually has.
Step 3, the ground noise, and the spike that is not ripple. The switch current jumps from zero to the 2.40 A peak in about 10 ns (AndwinPCB: rise times under 10 ns). A 10 mm trace adds 10 nH of parasitic inductance (AndwinPCB), and V = L·di/dt gives 2.40 V across it. A single via, roughly 1 nH, gives 0.24 V. A 5 nH stretch of ground return gives 1.20 V. Against the resistive drop, 4 squares of 1-oz plane at 500 µΩ per square is 2 mΩ, and 2.40 A through it is 4.8 mV, the inductive spike is 500 times the resistive drop, and it is 100 times the A7 core's 25 mV ripple budget. Put it against the victim: a 16-bit ADC on a 5 V reference has one least significant bit of 76 µV (CalcEngines). The 2.40 V spike is 31,579 LSBs. The 4.8 mV resistive drop is 63 LSBs. The 13 mV ripple is 171 LSBs. The 25 mV A7 limit is 329 LSBs. The analog readings are noisy because the ground under the ADC is moving by thousands of LSBs, ten thousand times a second, at the switching frequency. No amount of averaging in firmware fixes a ground that moves by 31,579 LSBs, averaging cuts noise as 1/√N, and the noise is not random, it is the switching frequency, and it is synchronous with the measurement.
Step 4, the input vs. the output, in harmonics. The input current is a square wave, 0 to 2.40 A and back, twice per cycle. The output current is a triangle, 2.0 A ± 0.40 A. The harmonics fall as 1/n for the square wave and 1/n² for the triangle (Vicor). At the 10th harmonic (5 MHz): the input still carries 0.240 A, the output 7.9 mA, the input is 30 times noisier. At the 60th harmonic (30 MHz, inside the FCC's radiated band): the input carries 0.040 A, the output 0.2 mA, 182 times noisier. The output filter is not the problem. The input is. Filter at the input, place the input capacitor within 5 mm of the IC, and keep the input loop small, because the input loop is the one that radiates.
Step 5, the loop area, and what the fix is worth. The radiated field from a current loop follows E = 1.32·A·I·F²/r (from Article #19: A in cm², F in MHz, r in metres, E in µV/m). The hot loop as drawn is 200 mm² (2 cm², a stated assumption), and at the 60th harmonic (30 MHz) it contributes 31.6 µV/m at 3 metres, under the FCC Class B limit of 100 µV/m at 30–88 MHz, on its own. The loop alone is not the whole failure: the switch node's capacitive coupling, the input cable (Article #19: below 1 GHz the radiator is usually a cable or an ungrounded unit), and the common-mode path add on top, and the field experience is that supplies laid out like this "radiated 15 dB over limits" (AndwinPCB). Shrink the loop from 200 mm² to 50 mm² and the field from the loop falls by a factor of 4, 12 dB, because the field is proportional to the area. Add the input filter, the sealed inductor (TI: a fully sealed inductor reduces the magnetic coupling to the measurement network), and the switch node kept under 20 mm², and the 15 dB excess is covered. The fix is not a shield. It is 150 mm² of copper that should not have been there, and a capacitor that should have been 5 mm from the IC.
The decision test
Five questions. If you can answer all five with a measurement from your own board, the rail is quiet. If you cannot, the layout is the problem.
- What is the ripple on the rail, measured at the output capacitor with a spring-tip probe, and what is the budget? 50–100 mV for most devices, 25 mV for a sensitive core rail. If the measurement was taken with a long probe ground, measure it again. The long ground is an antenna, and it will read several times the real noise.
- What is the ESR of the output capacitor at the switching frequency, and what is the capacitance at the working voltage? If the ESR term is bigger than the capacitance term, a bigger capacitor will not help. If the ceramic is 22 µF on the schematic and 3 µF at 24 V, the ripple calculation is wrong before it starts.
- How big is the hot loop, and how far is the input capacitor from the IC? The three grounds, input cap, device, output cap, as close together as possible, on a solid ground plane on the same layer as the components. The input capacitor within 5 mm of VIN and PGND. If the input capacitor's ground returns through the internal ground plane, the switching current flows through the board's ground.
- How big is the switch node, and what is next to it? Under 20 mm² for a design under 5 A, the inductor within 3 mm of the SW pin, and nothing analog next to it. The switch node is the primary source of EMI. If it is routed across the board, it is an antenna across the board.
- What is the ground noise at the ADC, in LSBs, not in millivolts? A 16-bit ADC on a 5 V reference has a 76 µV LSB. A 2.40 V spike is 31,579 LSBs. If the ground under the ADC moves by thousands of LSBs at the switching frequency, no firmware averaging will save the reading. The fix is the layout, not the filter.
Objections
"The schematic is correct, we simulated it." The simulation does not have a hot loop in it. It does not have 10 nH of trace inductance, or 500 µΩ per square of ground plane, or a ceramic capacitor that loses 85% of its capacitance at 24 V. The schematic is the circuit. The layout is the circuit plus the parasitics, and the parasitics are the noise. "When designing buck converters or boost regulators above 100 kHz, layout becomes the circuit, not just component arrangement."
"We added more output capacitance and the ripple didn't change." Because the ripple at the switching frequency is set by the ESR, not the capacitance, "the ESR of the capacitors is more important than their capacitance." A bigger capacitor with the same ESR gives the same ripple. A lower-ESR capacitor, or a smaller inductor ripple current, gives less. And if the capacitance you added is a ceramic at 24 V, you may have added 3 µF, not 22 µF.
"The board works on the bench and fails in the chamber." The bench is not the chamber. The bench has short cables, no LISN, and a probe ground that picks up several times the real noise from the air. The chamber measures what the board radiates, and what it radiates is the hot loop, the switch node, and the input cable. The board that "works" on the bench and fails in the chamber has a layout problem that the bench cannot see.
"We'll add a shield." A shield is a confession that the layout lost. It costs money, it costs weight, it traps heat, and it does not fix the ground noise under the ADC or the ripple on the rail. Shrink the loop by 150 mm² and the field falls 12 dB. Move the input capacitor to within 5 mm of the IC and the spike falls with it. The shield is the last resort, not the first.
"The motor doesn't care about the ripple." The motor doesn't. The sensor does, the ADC does, the BLE radio at 2.4 GHz does, and the chamber does. The EEVblog thread's original poster had the same question, 100 mV of ripple on a motor controller, and the answer was that it probably doesn't matter for the motor, but the board layout, grounding, and measurement still matter, because the same board will be the base for a power supply that does care. Fix the layout once, and the next board inherits it.
Sources
- EEPower, "Reducing Noise in an SMPS Through Layout and Filtering"
What we used
"Voltage ripple refers to the small AC voltage that sits on top of the regulated DC output. The goal is to minimize the amplitude of this voltage ripple, to present as clean of a DC source as possible to the rest of the circuit." "The threshold of ripple depends on what the supply is powering. 50mV to 100mV is likely an acceptable range for most devices, but some are more sensitive." "When running the A7 core at 900MHz, the VDD_ARM_CAP rail should be set to 1.275. However, it will work at 1.3V (25mV higher) or 1.25V (25mV lower). The acceptable voltage ripple on this rail is 25mV." "Electromagnetic interference (EMI) refers to unwanted electromagnetic emissions that can cause issues in otherwise functional circuits. The high frequency switching of the SMPS causes it to behave like a radiating antenna, emitting electrical signals." "The loop that the switching currents in an SMPS travel in is referred to as the 'hot loop.' We seek to minimize the size of this loop to limit the EMI being generated by the switching currents." "The hot loop for a buck converter circuit travels through the inductor, through the output capacitor, and back through the ground plane to the device ground and input capacitor ground. The simplest way to keep the hot loop as small as possible is to keep the three grounds (input cap, device, and output cap) as close together as possible. They should be connected with a solid ground plane on the same layer as the components, such that the path is as short as possible, and the violent switching currents avoid the internal PCB ground plane entirely." "It is also important to keep the switch node as small as possible. In a buck converter, this is the trace between the IC and the inductor, where currents are switched on and off at the switching frequency. This node is the primary source of EMI. The switch node behaves like a radiating antenna and by limiting the physical dimensions of the trace, the generated emissions can be limited."
- EEPower, "Optimize Switching Power Supply Layout by Minimizing Hot Loop PCB ESRs and ESLs"
What we used
"The hot loop of a switching-mode power converter is defined as the critical high frequency (HF) AC current loop formed by the HF capacitor and adjacent power FETs. It is the most critical part of the power stage PCB layout because it contains high dv/dt and di/dt noisy content." "For power converters, a hot loop PCB layout with minimum parasitic parameters can improve power efficiency, lower voltage ringing, and reduce electromagnetic interference (EMI)." "The switching actions of M1 and M2 cause HF di/dt and dv/dt noise. CIN provides a low impedance path to bypass the HF noisy content. However, parasitic impedance (ESRs, ESLs) exists within the components' packages and along the hot loop PCB traces. The high di/dt noise through ESLs causes HF ringing, furthermore, resulting in EMI. The energy stored in ESL is dissipated on ESRs, leading to extra power loss."
- Kueck (Linear Technology, AN139; via Analog / Power Electronics Technology, January 2013), "EMI starts off from high di/dt loops... To minimize EMI, we have to determine the 'hot' loop in the circuit and reduce its impact."
What we used
"Only in the green loop flows a fully switched AC current, switched from zero to IPEAK and back to zero. We refer to the green loop as a hot loop, because it has the highest AC and EMI energy." "To reduce EMI and improve functionality, you need to reduce the radiating effect of the green loop as much as possible. If we could reduce the PC-board area of the green loop to zero and buy an ideal input capacitor with zero impedance, the problem would be solved." "The supply wire as well as the load wire should not have high AC current content. So, we can focus our analysis from the input capacitor, CIN, which should source all relevant AC currents to the output capacitor, COUT, where any AC currents end."
- AndwinPCB, "Power Supply PCB Design, Layout Best Practices for Switching Regulators"
What we used
"Power supply failures account for 35% of PCB-related field returns, most from layout mistakes. When designing buck converters or boost regulators above 100 kHz, layout becomes the circuit, not just component arrangement." "I've debugged supplies that met electrical specs but radiated 15 dB over limits or exhibited 200 mV ripple instead of 20 mV." "Switching regulators create pulsed waveforms with rise times under 10 ns. At 500 kHz, fast edges make every trace a transmission line. A 10mm trace adds 10 nH parasitic inductance, generating 5V spikes when 10A switches in 10 ns, from V = L(di/dt)." "The hot loop, the high di/dt path where energy switches between input capacitor, high-side FET, low-side FET, and ground, operates at full switching current with nanosecond rise times. Minimize its area or accept radiated EMI that fails compliance." "Position input bulk capacitors within 5mm of the switching IC's VIN and PGND pins." "Every millimeter adds inductance converting switching current into voltage spikes. For 5A switching in 8ns, even 5nH generates a 3V spike from L(di/dt)." "Use ceramic capacitors with X7R or X5R dielectrics, never Y5V which loses 80% capacitance at rated voltage." "The switching node, connecting high-side FET, low-side FET, and inductor, carries full load current with voltage edges slewing at 5-20 V/ns. This node radiates broadband noise from tens to hundreds of MHz. Keep switching node copper under 20 mm² for <5A designs and under 50 mm² for 10A+. Use minimum copper area satisfying current density, typically 1 oz copper at 1A per 10 mil width for 10°C rise."
- EEVblog forum, "how to reduce buck regulator output ripple?" (thread read in full)
What we used
"One thing about buck converters is that they have much worse emi at the input than the output. The output current is triangular while the input is has discontinuous jumps. So pay particular attention to layout and filtering at the input." "For oscillations of the same frequency as the switching (400kHz), in this case the ESR of the capacitors is more important than their capacitance. What is the series resistance of capacitors? If you multiply that resistance by the variation in the inductor current, you will probably find the ripple voltage as solution." The thread's own case: a 100 mV+ ripple on an RAA211820 buck, 500 mA load, 18.3 V output, with "short big spikes", the replies diagnose layout, grounding, and measurement, not the circuit: "the board layout, grounding and how you do the measurement are very important"; "A 50V 10uF 0603 cap is like 1.5uF at 24V due to DC bias"; "Solder a 100uF electrolytic on top of the existing caps just as a quick test"; "First sanity-check the measurement: use a tip-and-barrel probe right at the output cap, those spikes often come from long probe grounds"; "fix layout and the hot loop first: put a few low-ESR ceramics (e.g., 2×10–22 µF) tight between VIN and PGND, same for VOUT and PGND, and keep the SW node as small as possible. Increase L so inductor ripple is ~20–40% of load current (if ΔIL is big, 100 mVpp is easy to hit). Add more output cap or a small 'damping' cap with a series 0.1–0.3 Ω to kill peaking. For spikes, an RC snubber from SW to GND (start ~1–3 nF with 2–5 Ω) helps a lot. Keep FB short and quiet; add a Cff (≈22–100 pF) across the top FB resistor." "Measuring the noise of DC-DC converters is surprisingly difficult. First thing to do is to remove any leads on the oscilloscope probe and use those small spring contacts to do the measurement. But the 1MOhm impedance of the probe means it will pick up EMC from the air, and I found that often to be several times the actual noise. Much better to measure this with a power rail probe, that has 50KOhm DC -50Ohm AC controlled impedance."
- AllPCB, "Practical EMC Design for Buck Converters"
What we used
"A buck converter has two main loops where current changes abruptly... Both currents, I1 and I2, are discontinuous, meaning they have steep rising and falling edges during switching. These edges have very short rise and fall times, resulting in a high rate of current change (dI/dt) and containing many high-frequency components." "In the loops described above, I1 and I2 share the path from the switch node → inductor → output capacitor → ground → Q2 source. When combined, I1 and I2 form a relatively smooth, continuous sawtooth waveform. Since this combined current lacks sharp edges with high dI/dt, it contains fewer high-frequency components." "From an EMI perspective, the shaded area A1 represents the loop with high dI/dt. This loop generates the most high-frequency components and is the most critical part to consider in the EMI design of a buck converter." "Due to the discontinuous nature of the buck converter's input current and the typically long power supply lines, the radiation from the input loop A3 can be [significant]." "Ideally, input and output capacitors present a very low impedance to the buck converter's switching currents. In practice, however, capacitors have ESR and ESL, which increase their impedance and cause additional high-frequency voltage drops."
- TI, Application Note SNVAA93, "The switcher's input current iQ1 is discontinuous resulting in fundamental and lower-order harmonics as well as high frequency noise current. Differential noise is caused by the opposite propagation paths between the positive and return power lines and is related to the fast di/dt of the input current."
What we used
"Besides, the input current loop can induce magnetic field that bypasses the EMI filter and couples to the LISN... it is critical to minimize the input current loop of the buck converter by strategically placing the input capacitor to reduce magnetic coupling and it is also beneficial to use a fully sealed inductor to reduce magnetic coupling."
- Vicor, "Your Power System Failed its Conducted EMI Test, Now What?"
What we used
"Input current through PWMSW rises and falls resembling a square wave as it is disconnected from the input each switching cycle." "Square wave, harm(n) α 1/n. Triangle wave, harm(n) α 1/n²." "Look at basic buck vs. boost converter: Input current through LSW rises and falls like a triangle wave, but it NEVER is disconnected from the input." "Ipp through LIN is typically 10% to 50% of the average current."
- EDN, "Use DC/DC buck converter features to optimize EMI in automotive designs"
What we used
"The most important practices for a buck converter are to: reduce the surface area of nodes with quickly changing voltages (high dv/dt), and reduce the area of current loops with quickly changing currents (high di/dt)." "DC/DC converters generate EMI from input ripple, electric and magnetic coupling to nearby circuits, and electromagnetic radiation. EMI can interfere with the AM/FM radio receiver and other sensitive equipment like the head unit or advanced driver assistance system (ADAS) sensors."
- RottenWifi, "How to Reduce DC/DC Converter Ground Noise"
What we used
"DC/DC converter ground noise is usually a voltage developed across real impedance, not noise generated by an abstract ground node. Fast switching currents flow through trace, via, package, capacitor, connector, and plane inductance, producing V = L·di/dt." "Two PCB locations can share the same schematic ground symbol while having different instantaneous voltages. A switching current pulse through copper creates resistive drop; through parasitic inductance it creates a much faster spike: V = L × di/dt." "A 1-ounce copper plane is approximately 500 µΩ per square, so a 1 A change can produce about 500 µV per square of resistive drop."
- EDN, "Reducing ground bounce in dc/dc-converter applications"
What we used
"'Ground bounce' is the amount that a ground return rises or falls relative to the system's 0V reference, and, in a dc/dc-switching converter, ground bounce can be many volts, often because of changing magnetic flux." "Generally, pc-board-ground-plane resistance is a less important source of ground bounce than magnetic-flux change. (The sheet resistance of 1-oz copper is about 500 µΩ/□, so a 1A change in current produces 500 µV/□ of bounce, a problem for thin, long, or daisy-chained grounds or precision electronics.)" "Therefore, the best way to reduce ground bounce in a switching dc/dc converter is to control changes in magnetic flux."
- Academy of EMC, "Why DC/DC Converters Fail EMC, and How to Fix It"
What we used
"Use a multilayer PCB with a solid ground (GND) plane." "Minimize the size of critical current branches." "Reduce coupling from the switching node to the feedback path and output." "Do not split or cut the ground plane beneath the DC/DC converter circuit. Instead, use a solid GND plane, which most effectively minimizes the loop areas of critical current paths." "The critical current branches are those with high di/dt [A/sec]. The branches containing the switch and the diode are always critical, because the current through the switch and diode can change from several amperes to 0 A in just a few nanoseconds."
- Power Electronics News, "PCB Layout for EMC, Part 2"
What we used
"If the noise matches the switching frequency of your switching converter that's already a pretty clear sign of who the guilty party is." "Switching converters are sources of noise. There's no avoiding this fact. And switchers emit radiofrequency noise." "The square waves, trapezoid waves and triangle waves that form most of those ripples have harmonics extending up to many times the switching frequency." "If you simply place vias into the ground plane at the negative connections for the loads of analog and noisy then you run a serious risk that the return currents will intermingle."
- TI SLVA477 / Altium / schemalyzer (buck converter sizing formulas, corroborated across three sources)
What we used
D = VOUT / (VIN × η); ΔIL = (VIN − VOUT) × D / (fSW × L); I(SW,max) = I(Load) + 0.5 × ΔIL; ripple ratio 20–40% of IOUT; L = VOUT × (VIN − VOUT) / (ΔIL × fS × VIN); output ripple ΔV ≈ ΔIL × ESR + ΔIL / (8 × fSW × C); COUT = ΔIL / (8 × fSW × ΔVOUT)
- CalcEngines, "Ground Loop Coupling Calculator"
What we used
"For a 16-bit converter on a 5 V reference, one least significant bit is 76 µV."
- Demo arithmetic verified in Python before writing: D = 0.2083; ΔIL = 0.792 A pk-pk (39.6% of IOUT
What we used
At the top of the 20–40% rule); IL,peak = 2.40 A; with 22 µH: ΔIL = 0.360 A, IL,peak = 2.18 A. Output ripple: ceramic 22 µF (ESR 5 mΩ, stated assumption): 4.0 mV ESR + 9.0 mV C = 13.0 mV; electrolytic 22 µF (ESR 100 mΩ, stated assumption): 79.2 mV ESR + 9.0 mV C = 88.2 mV; ceramic derated to 3 µF by DC bias: 4.0 mV + 66.0 mV = 69.9 mV. Input cap RMS current = 0.81 A. di/dt = 2.40 A / 10 ns = 2.4 × 10⁸ A/s; 10 nH (10 mm trace) → 2.40 V; 1 nH → 0.240 V; 5 nH → 1.20 V. Resistive: 4 squares × 500 µΩ = 2 mΩ → 4.8 mV at 2.40 A. LSBs (76 µV): 2.40 V = 31,579; 4.8 mV = 63; 13 mV = 171; 25 mV = 329. Harmonics: 10th (5 MHz) input 0.240 A vs. output 7.9 mA (30×); 60th (30 MHz) input 0.040 A vs. output 0.2 mA (182×); 100th (50 MHz) input 0.024 A vs. output 0.1 mA (303×). Radiated field (#19 formula, A in cm², F in MHz, r = 3 m): loop 200 mm² → 0.527 µV/m at the fundamental, 31.6 µV/m at the 60th harmonic vs. Class B 100 µV/m; loop 50 mm² → 0.132 µV/m and 7.9 µV/m; area 200 → 50 mm² = 4× field reduction = 12.0 dB