Menu

The Rail That Sags

The Rail That Sags

Notes · Power and batteriesNovember 19, 202622 min readSenova Systems Research

The processor wakes up, the rail droops, the board resets, the clock jitters, the radio loses packets, and the schematic is correct, because the problem is the impedance of the power delivery network, not the regulator.

The processor wakes up, the rail droops, the board resets, the clock jitters, the radio loses packets, and the schematic is correct. The regulator is the right part, the capacitors are the right values, the schematic has been checked three times, and the board still resets at bring-up. The problem is not the regulator.

The problem is the impedance of the power delivery network, the copper, the vias, and the capacitors between the regulator and the chip, and the schematic has no way of showing it, because the schematic has no inductance.

The simple idea first

Think of the power delivery network as a water supply, and the processor as a tap that opens and closes. The regulator is the water tower, it holds the pressure steady, but it is a long way from the tap, and it can only change its flow slowly. The decoupling capacitors are the tanks next to the tap, they hold a little water right where it is needed, and they can give it up fast.

The planes are the pipes, wide and flat, so the water can flow anywhere without much resistance. And the target impedance is the size of the pipe: if the pipe is too narrow, the pressure drops every time the tap opens, and the tap stutters.

That is the whole network, in one image. The regulator handles the slow changes, the bulk capacitors handle the medium ones, the small ceramics handle the fast ones, and the planes handle the fastest ones. Each one owns a frequency band, and the art of the design is making sure the bands hand off to each other without a gap, because a gap is a frequency where the rail rings, and the ring is the reset.

The target: one number, and the whole design follows from it

Power integrity comes down to one number and one shape: keep the PDN impedance below the target impedance, and keep the profile flat, no resonant peaks. (PCBSync) The target impedance is the ceiling your PDN impedance must stay under so the worst-case transient current produces an acceptable ripple. The formula is just Ohm's law: Ztarget = ΔV / ΔI, where ΔV is the allowed ripple and I_transient is the worst-case current step.

(PCBSync) Target impedance gives you a design limit for the rail so switching events do not create excessive ripple, ground bounce, jitter, or other behavior that shows up later as a signal integrity problem. (Flux)

The worked examples, from the sources: a 1.2 V rail with 5% ripple tolerance (60 mV) experiencing a 2 A transient requires a target impedance of 30 mΩ or lower across the operating band. (Flux) A 1.0 V core rail with a 5% budget (50 mV) and a 10 A transient needs 5 mΩ. (PCBSync) A 3.3 V rail with 5% tolerance (±165 mV) switching 8 A needs Ztarget = 165 mV / 8 A = 20.6 mΩ.

(Andwin) A processor core rail at 0.9 V, ±3% (±27 mV), 30 A transient: Ztarget = 0.9 × 0.03 / 30 = 0.9 mΩ, and the PDN impedance must remain below 0.9 milliohms from DC all the way up to the processor's maximum switching frequency. (AtlasPCB) And the article's own arithmetic on the same formula, labelled: a 1.0 V core rail with ±3% tolerance allows 30 mV, and a 2 A step gives 30 mV / 2 A = 15 mΩ.

And the number is not just small, it is small across a band. This impedance must stay below target from DC to the highest switching frequency of your load, typically 10 MHz for microcontrollers, 200 MHz for DDR4 memory, and 1 GHz for FPGAs. (Andwin) The PDN bandwidth requirement is set by the fastest current transient the load can produce: fmax = 0.5 / trise, a processor with a 1 ns current rise time needs 500 MHz.

(hilelectronic) The design goal is to maintain Z below Ztarget continuously from DC through fmax. Any frequency gap where PDN impedance exceeds Ztarget is a potential source of voltage droop, oscillation, or logic failure at the corresponding switching frequency. (hilelectronic)

The rule of thumb does not know any of these numbers. Power Distribution Network design goes beyond placing a 0.1 µF capacitor next to every power pin: it requires management of planar impedance, loop inductance, and transient current demands across a frequency spectrum from DC to several gigahertz. (Flux) The rule of thumb, one 100 nF ceramic on every power pin, one 10 µF bulk capacitor for each rail, is a guess, and the guess does not know the rail voltage, the ripple budget, or the load step.

The target impedance is the number that replaces the guess.

The rail: what the sag actually is

High-speed digital components fail when power delivery wavers. Voltage droops cause timing violations, and excess ripple induces clock jitter. (Flux) Poor PDN design guarantees erratic system behavior.

Surprisingly, power integrity problems are usually not caused by a lack of capacitance in the abstract, they emerge when the PDN presents an effective transfer impedance that is too high across the specific frequency range where the silicon demands transient current. When the PDN can't supply transient current fast enough, the local voltage droops, shrinks the operating margin, and leads directly to random resets and logic errors. (Flux) Every signal integrity problem has a power integrity root cause waiting to be discovered.

When your 10 Gbps SerDes link shows unexplained jitter, when your ADC's ENOB drops by two bits under load, when your processor throws intermittent errors during burst computation, the culprit is often a poorly designed Power Delivery Network. (AtlasPCB)

The sag has three parts, and the article that names them is the one that can fix them. The voltage drop depends on: resistive droop (V = I × R), from the DC resistance of planes, vias, and traces; inductive droop (V = L × di/dt), from connection inductance during fast transients; and resonance-induced overshoot, from PDN resonances excited by the current step. (AtlasPCB) And the sag comes in two events: the first droop (10–50 ns), limited by on-die capacitance and the nearest decoupling caps, the IC's internal capacitance absorbs the initial transient; and the second droop (100 ns–10 µs), limited by board-level decoupling and VRM response, typically the deeper droop event, the one that determines the minimum operating voltage.

(AtlasPCB)

The budget for the sag, from AtlasPCB's own droop-budget example for a 1.0 V core rail with ±5% tolerance (a 50 mV budget): VRM static accuracy ±10 mV; resistive droop 10 mV; first droop 10 mV; second droop 15 mV; margin 5 mV. (AtlasPCB) The budget adds to exactly 50 mV, and the sag that exceeds the budget is the sag that resets the board.

And the sag becomes the emissions. Power integrity issues such as impedance resonance peaks negatively influence the signal integrity behavior of boards, and a noisy PDN can easily become a strong parasitic EMI antenna. If a ground plane bounces due to high loop inductance in the return path, the reference voltage for the entire logic bus shifts.

This translates directly into jitter, timing violations, and radiated electromagnetic interference (EMI) failures. (Flux) The rail that sags is the rail that radiates, and the board that fails EMC is often the board whose PDN was never designed, only populated.

The capacitor

A real capacitor is not just capacitance, it is an RLC circuit: ESR (equivalent series resistance), which dominates at the self-resonant frequency; ESL (equivalent series inductance), which dominates above SRF, so the capacitor becomes an inductor; and capacitance, which dominates below SRF and provides decreasing impedance with frequency. (AtlasPCB) Below the SRF, a capacitor acts like a capacitor. Above it, it acts like an inductor.

A single capacitor cannot achieve low impedance across the full frequency range. (Flux) The self-resonant frequency, where impedance is minimum: SRF = 1 / (2π × √(L × C)). (AtlasPCB)

The worked examples: a typical 100 nF 0402 MLCC with 0.5 nH ESL has SRF = 1 / (2π × √(0.5×10⁻⁹ × 100×10⁻⁹)) = 22.5 MHz, and above 22.5 MHz, this capacitor behaves as an inductor, not a capacitor. (AtlasPCB) A 100 nF capacitor in 0805 resonates at 14 MHz versus 19 MHz in 0402, an 0805 package has approximately 1.2 nH of equivalent series inductance versus 0.7 nH for a 0402. (Flux) The SRF table, from AtlasPCB's own page: 100 µF 1210, ~500 kHz, effective band 100 kHz–2 MHz; 10 µF 0805, ~2 MHz, 500 kHz–5 MHz; 1 µF 0402, ~7 MHz, 2–20 MHz; 100 nF 0402, ~22 MHz, 5–50 MHz; 10 nF 0201, ~70 MHz, 20–150 MHz; 1 nF 0201, ~220 MHz, 50–500 MHz.

(AtlasPCB) And a 10 µF capacitor has an SRF of approximately 500 kHz, above 1 MHz it behaves inductively, demonstrating steadily rising impedance, and it cannot handle 100+ MHz switching transients. (Flux)

No single capacitor value covers the entire operating frequency band. Effective decoupling requires multiple capacitor values with staggered self-resonant frequencies to maintain a flat impedance profile and avoid anti-resonance peaks. (Flux) The combinations, from Flux: 0.01–0.1 µF placed immediately next to power pins, for the highest frequency transients; 1–4.7 µF placed slightly farther away, for midrange frequencies; 10–47 µF bulk near the power entry point or VRM, for slow, large load steps.

Choose at least 3 different capacitor values to avoid high resonant peaks when paralleling ceramics. (Flux) Andwin's version: the solution is multi-decade capacitor spacing, typical combinations are 10 µF, 1 µF, and 0.1 µF in ratios of 10:1, and modern FPGA designs targeting sub-5 mΩ impedance use four decades: 100 µF (VRM side), 10 µF (regional placement), 1 µF (IC side), 0.1 µF (per power ball). (Andwin)

The anti-resonance: the peak between the values

There is also a less-obvious trap: anti-resonance. Placing one large capacitor in parallel with one small capacitor creates an anti-resonance, a frequency at which the inductance of the larger cap resonates with the capacitance of the smaller cap. This anti-resonance produces an impedance peak that can be higher than either capacitor alone would produce.

(Flux) When two capacitors of different values are connected in parallel, they form an LC circuit at the frequency where the larger cap becomes inductive and the smaller cap is still capacitive. This creates an anti-resonance peak, a frequency where PDN impedance spikes dramatically. (AtlasPCB)

The worked example, from PCBSync: the anti-resonance sits at f_anti = 1 / (2π√(L_big × C_small)). Run it on the classic pairing, a 10 µF 0805 at 0.9 nH beside a 100 nF 0402, and the peak lands at 16.8 MHz. A 10 µF 0805 at 0.9 nH with a 100 nF 0402 at 0.5 nH puts the anti-resonance at 16.8 MHz while the 100 nF's SRF is 22.5 MHz.

At 16.8 MHz the 100 nF is still firmly capacitive, the tank is real, and so is the peak. (PCBSync) The decade rule, 10 µF, 1 µF, 100 nF, 10 nF, puts an impedance peak between every adjacent pair. (PCBSync) And the peak is not a curiosity: if a resonance peak appears at 50 MHz with an impedance of 100 mΩ, far above the 5 mΩ target, it indicates a need for design adjustments, such as adding decoupling capacitors.

(allpcb)

The mitigations, from AtlasPCB: use many different values, more values, smaller anti-resonance peaks; add lossy capacitors, higher ESR damps resonances, and some designers intentionally add resistance; rely on plane capacitance, at high frequencies, the power/ground plane pair provides broadband capacitance with inherent damping; and simulate, don't guess, PDN simulation tools reveal anti-resonance locations. (AtlasPCB) Target impedance is a line to stay under at every frequency, not a single number to hit.

You can size each capacitor correctly and still fail, because between two capacitor values the network can resonate and spike above the target. (PCBSync)

The frequency domains: who owns which band

Each element in the PDN owns a specific frequency band. The art of PDN design is ensuring smooth impedance handoff between these elements with no resonant peaks that exceed the target. (AtlasPCB) The table, from AtlasPCB: VRM (regulator), DC–10 kHz, set by the regulation loop; bulk electrolytic caps, 1 kHz–100 kHz, 10–100 mΩ; bulk MLCC (47–100 µF), 10 kHz–2 MHz, 2–20 mΩ; decoupling MLCC (100 nF–1 µF), 500 kHz–100 MHz, 1–10 mΩ; high-frequency MLCC (1–10 nF), 10 MHz–500 MHz, 1–5 mΩ; power/ground plane pair, 100 MHz–2 GHz, 0.5–5 mΩ; on-die capacitance, above 1 GHz, IC-dependent.

(AtlasPCB) Flux's version of the same table: VRM, DC to ~100 kHz; bulk capacitors, 100 kHz to ~1 MHz; decoupling capacitors, 1 MHz to ~200 MHz; power/ground planes, 100 MHz to ~1 GHz. (Flux) And above a few hundred MHz the board cannot help, that job belongs to package and on-die capacitance. (PCBSync)

The VRM owns the bottom of the band, and only the bottom. PI events can only be managed within the VRM's control loop bandwidth, which is typically in the 100 kHz range. Above that, bulk and decoupling capacitors carry the load until plane capacitance takes over.

(Flux) The regulator is the water tower, it holds the pressure steady, but it is a long way from the tap, and it can only change its flow slowly. When millions of transistors switch simultaneously, the VRM is electrically too far away to respond in time. The local decoupling capacitor supplies that instantaneous current demand.

(Flux)

The planes own the top of the band, and they own it because they are everywhere. A parallel-plate capacitor is formed by every adjacent power/ground plane pair: C_plane = ε₀ × εr × Area / thickness. For a 100 mm × 80 mm board, 0.1 mm (4 mil) dielectric, εr 4.2: C_plane = 8.854×10⁻¹² × 4.2 × 8×10⁻³ / 0.1×10⁻³ = 3.0 nF.

This 3 nF of plane capacitance is distributed across the entire board area with very low inductance, it is the most effective high-frequency PDN element. (AtlasPCB) Inter-plane capacitance is created by placing the power and ground planes very close to each other, typically less than 3 mils apart. Reducing the spacing between a power plane and ground plane from 10 mils to 5 mils can double the inter-plane capacitance, significantly lowering impedance at frequencies above 100 MHz.

(Flux) To maximize plane-pair capacitance: minimize dielectric spacing, use thin prepreg (2–4 mil) between power/ground pairs; maximize plane area, avoid splits and cutouts in the power region under ICs; place power/ground pairs adjacent, never separate them with signal layers. (AtlasPCB)

The loop: the millimetres are the impedance

Below ~50 MHz, placement does not matter much. At frequencies below ~50 MHz, the wavelength in FR-4 exceeds the board dimension, the entire power plane acts as a single equipotential node, and capacitor location has minimal impact, so bulk capacitors can be placed anywhere convenient. Above 50 MHz, connection inductance between the capacitor and the IC dominates.

(AtlasPCB) And the inductance is not small: a single via is ~0.5–1.0 nH; a trace is ~0.7–1.0 nH per millimetre; a via pair (signal + return) is ~0.3–0.7 nH. This is why high-frequency decoupling must be placed within 2 mm of IC power pins. (AtlasPCB) Every millimeter of trace between a capacitor and an IC power pin adds approximately 1 nH of inductance, shifting the effective self-resonant frequency downward and degrading high-frequency decoupling.

(Flux)

At 200 MHz, 1 nH of inductance presents Z=2π×200×10⁶×1×10⁻⁹=1.26 Ω, 1,000× higher than a typical target impedance\text{At 200 MHz, 1 nH of inductance presents Z} = \text{2π} \times \text{200×10⁶} \times \text{1×10⁻⁹} = \text{1.26 Ω, 1,000× higher than a typical target impedance}

The placement rules, from AtlasPCB: place 1–10 nF caps within 1–2 mm of IC power/ground pin pairs; place 100 nF caps within 5 mm of the IC; place 1–10 µF caps within 10–15 mm; place bulk caps (47–100 µF) anywhere on the board. (AtlasPCB) And the via rules: use multiple vias per cap pad (2 vias per pad for 0402, 3–4 for 0805+); connect vias directly to the pad, no trace between pad and via; use via-in-pad when possible; route the power via and ground via as close together as possible to minimize loop inductance.

(AtlasPCB) A standard 0.3 mm drill via in a 1.6 mm board contributes approximately 0.5–1 nH of inductance, comparable to or exceeding the package inductance of a 0201 capacitor. Via optimization (shorter vias, larger via diameter, multiple vias per pad) becomes as important as capacitor selection. (Flux) And the check from Article #19, which still holds: 100 nF per IC power pin, target under 2 mm from the pin, return via within 1 mm of the cap ground pad.

The mistakes, from AtlasPCB's own list: insufficient via connections, a single via on each cap pad adds ~1 nH, negating the capacitor's effectiveness above 50 MHz; plane splits under ICs, creates return current discontinuities and eliminates plane capacitance exactly where it is needed; excessive trace between cap and via, even 2 mm of trace adds significant inductance at high frequencies; wrong capacitor values, using only 100 nF (the "classic" bypass cap) leaves gaps below 5 MHz and above 50 MHz; ignoring ESL specifications, two 100 nF caps with different ESL provide different effective frequency ranges.

(AtlasPCB) And from Flux: "Swiss-cheese" power planes, routing dense via arrays through internal power and ground planes punches continuous holes in the copper, restricting current flow and increasing local inductance precisely where high-speed signals need a solid return path; splitting power planes without precaution; relying on a single bulk capacitor for high-frequency decoupling; routing signals across power plane splits; ignoring package size on decoupling capacitors; thermal relief spokes on high-current vias.

(Flux)

To hold 5 mΩ at 100 MHz, the total loop inductance has to stay under 8 picohenries. A mounted 0402 with its vias is somewhere around 1.5–2 nH. That is 250 capacitors in parallel, which is why a large FPGA board genuinely carries two hundred decaps.

(PCBSync) The impedance of an inductor is 2πfL: at 1 MHz you can afford 796 pH to hold 5 mΩ; at 10 MHz, 80 pH; at 100 MHz, 8 pH. (PCBSync) The number is not a suggestion. It is the whole design, in one line.

The hub board, and the 0.36 V sag

Example Labs is a made-up company we use for worked examples.

The Example Labs field sensor hub, 100 × 60 mm, 4-layer, 120 nets, 114 test points (#32), STM32H7 at 48 MHz, BLE radio at 2.4 GHz (#12), sensor on a 3.0 V rail with a 2.7 V minimum (#41). The demo is the article's own worked example, labelled, built on the cluster's figures and the verified sources above.

Step 1, the target. The sensor rail is 3.0 V, and the tolerance is ±5%, a budget of 3.0 × 0.05 = 150 mV. The load step, from Article #51's labelled assumption, is the hub's peak draw of 100 mA. The twin, from Article #41, predicted a sag band of 0.30–0.40 V and measured a sag of 0.36 V, and the rail fell to 3.0 − 0.36 = 2.64 V, 0.06 V below the sensor's 2.7 V minimum. The sag is 0.36 / 0.15 = 2.4× the budget, and the whole predicted band, 0.30–0.40 V, is 2.0× to 2.67× the budget. The budget is not the problem. The PDN is.

The target impedance is Ztarget=150 mV/0.1 A=1.5 Ω\text{The target impedance is Ztarget} = \text{150 mV} / \text{0.1 A} = \text{1.5 Ω}

Step 2, the impedance. The sag is Ohm's law in the time domain: ΔV = ΔI × Z. The measured impedance is 0.36 V / 0.1 A = 3.6 Ω, 2.4× over the 1.5 Ω target, 2.1 Ω above it. The regulator cannot help: the VRM's control loop bandwidth is typically in the 100 kHz range (Flux), and the step is at 48 MHz, 480× above the bandwidth where the regulator can respond. Above 100 kHz, the bulk and decoupling capacitors carry the load until plane capacitance takes over. (Flux) The regulator is the water tower, and the tap is opening 480 times faster than the tower can answer.

Step 3, the loop. The 100 nF capacitor sits 10 mm from the pin, on a single via. The trace is 0.7–1.0 nH per millimetre (AtlasPCB), 10 mm is 7–10 nH, and the single via is 0.5–1.0 nH. Take 8 nH for the loop, the article's own worked example, labelled. The capacitor's own ESL, 0.5 nH, adds 2π × 48×10⁶ × 0.5×10⁻⁹ = 0.15 Ω. The loop is most of the 3.6 Ω, and the loop is the millimetres, not the capacitor. Move the capacitor to within 2 mm of the pin, on a via pair: the loop falls to ~2 nH, and the reactance at 48 MHz falls to 2π × 48×10⁶ × 2×10⁻⁹ = 0.60 Ω. And the plane pair: the hub board is 100 × 60 mm, 0.1 mm dielectric, εr 4.2, C_plane = 8.854×10⁻¹² × 4.2 × 6×10⁻³ / 0.1×10⁻³ = 2.23 nF, the article's own arithmetic on AtlasPCB's formula (AtlasPCB's own 100 × 80 mm example is 3.0 nF). At 48 MHz the plane pair's reactance is 1 / (2π × 48×10⁶ × 2.23×10⁻⁹) = 1.49 Ω, right at the target, but the spreading inductance means the far parts of the plane cannot deliver current in the first 10–50 ns, which is why the nearest capacitor is the first responder and the plane is the second. The first droop (10–50 ns) is limited by on-die capacitance and the nearest decoupling caps; the second droop (100 ns–10 µs) is limited by board-level decoupling and VRM response. (AtlasPCB)

At 48 MHz, the loop’s reactance is X=2π×48×10⁶×8×10⁻⁹=2.41 Ω, already 1.6× the 1.5 Ω target, before the plane spreading inductance is counted\text{At 48 MHz, the loop's reactance is X} = \text{2π} \times \text{48×10⁶} \times \text{8×10⁻⁹} = \text{2.41 Ω, already 1.6× the 1.5 Ω target, before the plane spreading inductance is counted}

Step 4, the anti-resonance. The board carries a 10 µF 0805 bulk capacitor (0.9 nH ESL) beside the 100 nF 0402. The anti-resonance sits at f_anti = 1 / (2π√(L_big × C_small)) = 1 / (2π√(0.9×10⁻⁹ × 100×10⁻⁹)) = 16.8 MHz (PCBSync's own calculation, reproduced exactly). At 16.8 MHz the 100 nF is still firmly capacitive, its SRF is 22.5 MHz, and the 10 µF is inductive, so the tank is real, and so is the peak. (PCBSync) If the peak rises above the 1.5 Ω target, the rail rings at 16.8 MHz every time the load switches, and the ring is a sag that repeats, not a sag that happens once. The fix: add a 1 µF 0402 between them (SRF ~7 MHz, effective band 2–20 MHz, AtlasPCB), so the values hand off instead of peaking, more values, smaller peaks (AtlasPCB), or add lossy capacitance, whose higher ESR damps the resonance. (AtlasPCB)

Step 5, the ripple, and the fix. The buck that feeds the rail, from Article #44: 24 V → 5 V, 2 A, 500 kHz, L = 10 µH, D = 0.2083, ΔIL = 0.792 A, and the ripple is 13.0 mV with a ceramic output capacitor against 88.2 mV with an electrolytic. The ripple the PDN carries is set by the capacitor's ESR, not its capacitance, and the aluminium electrolytic's ESR is 50–200 mΩ against the polymer's 5–15 mΩ (Andwin). And the spike: 10 nH at 2.4 × 10⁸ A/s = 2.40 V (#44), the same V = L × di/dt that makes the rail sag, seen at the switching edge. The fix is the checklist, from AtlasPCB: target impedance calculated for each power rail; stackup includes at least one thin-core power/ground pair; 1–10 nF caps within 2 mm of every high-speed IC power pin; 100 nF caps within 5 mm of every IC, one per power pin minimum; 10–100 µF bulk caps near the VRM output, 1 per 5 A of load current; no power plane splits under ICs; via-in-pad or adjacent via for all decoupling caps; power plane pour connected with multiple vias at each junction; ferrite beads avoided on high-current paths; PDN simulation performed at critical frequencies. (AtlasPCB) And the cost of finding it late: the escalation from Article #40, 1× at the schematic, 5× at the prototype, 20× at the pilot, 50× at mass production, 100×+ in the field, and the field failure from Article #40, 4 of 10 hubs in days 19–23, 40%.

The decision test

Five questions. If the answer to any one of them is no, the PDN is a guess, and the decision is to measure first, and to design the network the measurement supports.

  1. What is the target impedance, in milliohms, for each rail, and what is the load step that sets it? Ztarget = ΔV / ΔI, where ΔV is the maximum allowable voltage ripple and I_transient is the maximum transient current. (Flux) If the target is not written down, the capacitors are a guess, and the rule of thumb does not know the rail voltage, the ripple budget, or the load step.
  2. Does the impedance stay below target at every frequency up to the knee frequency, or does it peak somewhere in between? Target impedance is a line to stay under at every frequency, not a single number to hit. You can size each capacitor correctly and still fail, because between two capacitor values the network can resonate and spike above the target. (PCBSync) The knee frequency is f_knee ≈ 0.35 / t_rise, there is little energy above it, so driving impedance down at 1 GHz is pointless. (PCBSync)
  3. Is there an anti-resonance peak between any two capacitor values, and where does it sit? The anti-resonance sits at f_anti = 1 / (2π√(L_big × C_small)), and the classic pairing, a 10 µF 0805 beside a 100 nF 0402, puts the peak at 16.8 MHz. (PCBSync) If the peak is above the target, the rail rings at that frequency every time the load switches.
  4. How far is the nearest capacitor from the pin, in millimetres, and how many vias does it have? At 200 MHz, 1 nH of inductance presents 1.26 Ω, 1,000× higher than a typical target impedance. This is why high-frequency decoupling must be placed within 2 mm of IC power pins. (AtlasPCB) A single via on each cap pad adds ~1 nH, negating the capacitor's effectiveness above 50 MHz. (AtlasPCB)
  5. Is the plane pair thin enough, and is it unbroken under the IC? Inter-plane capacitance is created by placing the power and ground planes very close to each other, typically less than 3 mils apart. Reducing the spacing from 10 mils to 5 mils can double the inter-plane capacitance. (Flux) Plane splits under ICs create return current discontinuities and eliminate plane capacitance exactly where it is needed. (AtlasPCB)

Objections

"The schematic is correct, the regulator is the right part, so the rail should be fine." The schematic has no inductance. Every trace, via, and copper pour adds parasitic inductance or capacitance that shapes the overall impedance profile. (Flux) The regulator is the water tower, it holds the pressure steady, but it can only change its flow slowly, and its control loop bandwidth is typically in the 100 kHz range. Above that, bulk and decoupling capacitors carry the load until plane capacitance takes over. (Flux) The regulator is the right part for the slow changes. The sag is a fast change, and the fast change is the network's job.

"We have enough capacitance, there are a hundred capacitors on the board." Power integrity problems are usually not caused by a lack of capacitance in the abstract, they emerge when the PDN presents an effective transfer impedance that is too high across the specific frequency range where the silicon demands transient current. (Flux) A hundred 100 nF capacitors is 10 µF of capacitance, and if they are all 10 mm from the pins, on single vias, the loop inductance is still 8 nH each, and the rail still sags. The count is not the design. The impedance is.

"One 100 nF per power pin is the standard, it has always worked." Using only 100 nF (the "classic" bypass cap) leaves gaps below 5 MHz and above 50 MHz. (AtlasPCB) A 10 µF capacitor has an SRF of approximately 500 kHz, above 1 MHz it behaves inductively, and a 100 nF 0402 has an SRF of 22.5 MHz, above which it behaves as an inductor. (Flux; AtlasPCB) The standard worked when the rails were 5 V and the clocks were 10 MHz. The hub's rail is 3.0 V with a 150 mV budget, and the clock is 48 MHz, and the standard does not know either number.

"We can just add a bigger bulk capacitor." A larger capacitor is not always better. Placing one large capacitor in parallel with one small capacitor creates an anti-resonance, a frequency at which the inductance of the larger cap resonates with the capacitance of the smaller cap, and the impedance peak can be higher than either capacitor alone would produce. (Flux) The bigger capacitor owns the low frequencies. The peak it creates with the small capacitor owns the middle. And the middle is where the load switches.

"Simulation is overkill, we will just measure it at bring-up." Analyzing a circuit layout for PDN impedance can hardly be done with a pen and paper, and PCB CAD tools cannot handle target impedance issues simply by defining a design rule. Advanced engineering tools like numerical PI solvers are needed. (Flux) And the measurement at bring-up is the expensive one: the escalation from Article #40, 1× at the schematic, 5× at the prototype, 20× at the pilot, 50× at mass production, 100×+ in the field. The simulation is the cheap measurement. The bring-up is the expensive one.

Sources

  1. Flux, "PDN Design in PCBs: Power Distribution Network Best Practices (2026)" (opened in full)
    What we used

    "High-speed digital components fail when power delivery wavers. Voltage droops cause timing violations, and excess ripple induces clock jitter." "Power Distribution Network (PDN) design for PCBs goes beyond placing a 0.1 µF capacitor next to every power pin: it requires management of planar impedance, loop inductance, and transient current demands across a frequency spectrum from DC to several gigahertz." "A PCB PDN is a broadband, frequency-dependent impedance network spanning the VRM to the IC die, composed of regulators, planes, vias, capacitors, and package parasitics." "Effective decoupling requires multiple capacitor values with staggered self-resonant frequencies to maintain a flat impedance profile and avoid anti-resonance peaks." "Target impedance should remain below specification across the full operating bandwidth through tight plane coupling, minimal loop inductance, and optimized via geometry." "The primary function of the PDN is to provide a low-impedance path from the Voltage Regulator Module (VRM) to the IC die. When an IC transitions from idle to maximum processing load, it demands an instantaneous surge of current. The PDN must supply that current without allowing the voltage rail to collapse, meaning it cannot let the localized voltage dip below the silicon's minimum logic threshold, which would cause transistors to starve and misinterpret binary states." "Poor PDN design in a PCB guarantees erratic system behavior. Surprisingly, power integrity problems in a high-speed PCB are usually not caused by a lack of capacitance in the abstract, they emerge when the PDN presents an effective transfer impedance that is too high across the specific frequency range where the silicon demands transient current. When the PDN can't supply transient current fast enough, the local voltage droops, shrinks the operating margin, and leads directly to random resets and logic errors." "Power integrity issues such as impedance resonance peaks negatively influence the signal integrity behavior of boards, and a noisy PDN can easily become a strong parasitic EMI antenna. If a ground plane bounces due to high loop inductance in the return path, the reference voltage for the entire logic bus shifts. This translates directly into jitter, timing violations, and radiated electromagnetic interference (EMI) failures." "Target impedance gives you a design limit for the rail so switching events do not create excessive ripple, ground bounce, jitter, or other behavior that shows up later as a signal integrity problem." Frequency table: VRM DC to ~100 kHz; bulk capacitors 100 kHz to ~1 MHz; decoupling capacitors 1 MHz to ~200 MHz; power/ground planes 100 MHz to ~1 GHz. "PI events can only be managed within the VRM's control loop bandwidth, which is typically in the 100 kHz range. Above that, bulk and decoupling capacitors carry the load until plane capacitance takes over." "Inter-plane capacitance is created by placing the power and ground planes very close to each other, typically less than 3 mils apart. Reducing the spacing between a power plane and ground plane from 10 mils to 5 mils can double the inter-plane capacitance, significantly lowering impedance at frequencies above 100 MHz." "Every millimeter of trace between a capacitor and an IC power pin adds approximately 1 nH of inductance, shifting the effective self-resonant frequency (SRF) downward and degrading high-frequency decoupling." "No single capacitor value covers the entire operating frequency band. Below the SRF, a capacitor acts like a capacitor. Above it, it acts like an inductor." "There's also a less-obvious trap: anti-resonance. A common misconception is that a larger capacitor is always better. In reality, placing one large capacitor in parallel with one small capacitor creates an anti-resonance, a frequency at which the inductance of the larger cap resonates with the capacitance of the smaller cap. This anti-resonance produces an impedance peak that can be higher than either capacitor alone would produce." Capacitor combinations: 0.01–0.1 µF immediately next to power pins; 1–4.7 µF slightly farther away; 10–47 µF bulk near the VRM. "Choose at least 3 different capacitor values to avoid high resonant peaks when paralleling ceramics." "Calculate your target impedance first. Target impedance is calculated using the formula Z_target = ΔV / I_transient, where ΔV is the maximum allowable voltage ripple and I_transient is the maximum transient current. For example, a 1.2V rail with 5% ripple tolerance (60 mV) experiencing a 2A transient requires a target impedance of 30 mΩ or lower across the operating band." "Use via-in-pad for capacitors wherever possible. A standard 0.3 mm drill via in a 1.6 mm board contributes approximately 0.5–1 nH of inductance, comparable to or exceeding the package inductance of a 0201 capacitor." "A 10 µF capacitor has an SRF of approximately 500 kHz; above 1 MHz, it behaves inductively, demonstrating steadily rising impedance. It cannot handle 100+ MHz switching transients." "A 0805 package has approximately 1.2 nH of equivalent series inductance (ESL) versus 0.7 nH for a 0402. A 100 nF capacitor in 0805 resonates at 14 MHz versus 19 MHz in 0402. Use the smallest package that fits for high-frequency positions." "Use simulation early: Analyzing a circuit layout for PDN impedance can hardly be done with a pen and paper, and PCB CAD tools cannot handle target impedance issues simply by defining a design rule. Advanced engineering tools like numerical PI solvers are needed."

  2. AtlasPCB, "PCB Power Integrity Design" (opened in full)
    What we used

    "Every signal integrity problem has a power integrity root cause waiting to be discovered. When your 10 Gbps SerDes link shows unexplained jitter, when your ADC's ENOB drops by two bits under load, when your processor throws intermittent errors during burst computation, the culprit is often a poorly designed Power Delivery Network (PDN)." "Core voltages keep dropping, Modern processors run at 0.7-1.0V, leaving almost no noise margin. Transient currents keep rising, A modern FPGA can demand 50A+ step-load changes in nanoseconds. Switching frequencies keep increasing, GHz clock rates mean femtosecond-scale timing margins." "Ztarget = Vdd × Ripple_Tolerance / I_transient_max. For a processor core rail: Vdd = 0.9V, allowable ripple = ±3% (±27mV), maximum transient current = 30A. Ztarget = 0.9 × 0.03 / 30 = 0.9 mΩ." "This means the PDN impedance, as seen from the IC, must remain below 0.9 milliohms from DC all the way up to the processor's maximum switching frequency." Frequency domains: VRM DC–10 kHz; bulk electrolytic 1 kHz–100 kHz, 10–100 mΩ; bulk MLCC 47–100 µF 10 kHz–2 MHz, 2–20 mΩ; decoupling MLCC 100 nF–1 µF 500 kHz–100 MHz, 1–10 mΩ; high-frequency MLCC 1–10 nF 10 MHz–500 MHz, 1–5 mΩ; power/ground plane pair 100 MHz–2 GHz, 0.5–5 mΩ; on-die capacitance above 1 GHz. "The art of PDN design is ensuring smooth impedance handoff between these elements with no resonant peaks (anti-resonances) that exceed the target." "A real capacitor is not just capacitance, it's an RLC circuit: ESR dominates at the self-resonant frequency; ESL dominates above SRF, capacitor becomes an inductor; capacitance dominates below SRF." "SRF = 1 / (2π × √(L × C)). For a typical 100nF 0402 MLCC with 0.5nH ESL: SRF = 1 / (2π × √(0.5×10⁻⁹ × 100×10⁻⁹)) = 22.5 MHz. Above 22.5 MHz, this capacitor behaves as an inductor, not a capacitor." Capacitor value table: 100 µF 1210 ~500 kHz, 100 kHz–2 MHz; 10 µF 0805 ~2 MHz, 500 kHz–5 MHz; 1 µF 0402 ~7 MHz, 2–20 MHz; 100 nF 0402 ~22 MHz, 5–50 MHz; 10 nF 0201 ~70 MHz, 20–150 MHz; 1 nF 0201 ~220 MHz, 50–500 MHz. "When two capacitors of different values are connected in parallel, they form an LC circuit at the frequency where the larger cap becomes inductive and the smaller cap is still capacitive. This creates an anti-resonance peak, a frequency where PDN impedance spikes dramatically." Mitigations: many different values; lossy capacitors (higher ESR damps resonances); plane capacitance; simulate, don't guess. "Below ~50 MHz, the wavelength in FR-4 (λ = c / (f × √εr) ≈ 150mm / f_GHz) exceeds the board dimension. The entire power plane acts as a single equipotential node, so capacitor location has minimal impact." "Above 50 MHz, connection inductance between the capacitor and the IC dominates. The inductance of a via + short trace is approximately: single via ~0.5–1.0 nH; trace (per mm) ~0.7–1.0 nH/mm; via pair (signal + return) ~0.3–0.7 nH. At 200 MHz, 1 nH of inductance presents: Z = 2π × 200×10⁶ × 1×10⁻⁹ = 1.26 Ω. That's 1,000× higher than a typical target impedance! This is why high-frequency decoupling must be placed within 2mm of IC power pins." Placement rules: 1–10 nF within 1–2 mm; 100 nF within 5 mm; 1–10 µF within 10–15 mm; bulk 47–100 µF anywhere. Via optimization: multiple vias per pad (2 per pad for 0402, 3–4 for 0805+); vias directly to pad; via-in-pad; power and ground vias close together. "C_plane = ε₀ × εr × Area / thickness. For a typical configuration: Area 100mm × 80mm = 8,000 mm², dielectric thickness 0.1mm (4 mil), εr 4.2: C_plane = 3.0 nF. This 3nF of plane capacitance is distributed across the entire board area with very low inductance, it's the most effective high-frequency PDN element." "The voltage drop depends on: resistive droop (V = I × R); inductive droop (V = L × di/dt); resonance-induced overshoot." "First droop (10–50 ns): limited by on-die capacitance and nearest decoupling caps. Second droop (100 ns–10 µs): limited by board-level decoupling and VRM response. This is typically the deeper droop event that determines minimum operating voltage." Droop budget for a 1.0 V core rail with ±5% tolerance (50 mV): VRM static accuracy ±10 mV; resistive droop 10 mV; first droop 10 mV; second droop 15 mV; margin 5 mV. Checklist: target impedance per rail; thin-core power/ground pair; 1–10 nF within 2 mm; 100 nF within 5 mm, one per power pin minimum; 10–100 µF bulk near VRM, 1 per 5 A of load current; no plane splits under ICs; via-in-pad; ferrite beads avoided on high-current paths; PDN simulation at critical frequencies. Mistakes: single via per pad adds ~1 nH, negating effectiveness above 50 MHz; plane splits under ICs; even 2 mm of trace adds significant inductance; only 100 nF leaves gaps below 5 MHz and above 50 MHz; ignoring ESL

  3. Andwin, "Power Integrity in PCB Design: PDN Impedance and Decoupling Strategy"
    What we used

    "Achieving target PDN impedance below 10% of load impedance across 1MHz-1GHz prevents voltage droops exceeding 5% during transient switching, the threshold where FPGA bit errors, microcontroller resets, and RF amplifier spurious emissions begin appearing in production. For a 1.0V core rail switching 10A in 50ns, this requires PDN impedance below 5mΩ." "The target impedance calculation is straightforward: Ztarget = ΔV / ΔI, where ΔV is your maximum allowable voltage ripple and ΔI is the peak current transient. For a 3.3V rail with 5% tolerance (±165mV) switching 8A, you need Ztarget = 165mV / 8A = 20.6mΩ." "This impedance must stay below target from DC to the highest switching frequency of your load, typically 10MHz for microcontrollers, 200MHz for DDR4 memory, and 1GHz for FPGAs." "Aluminum electrolytics have high ESR (50mΩ-200mΩ) but large capacitance, making them effective below 100kHz. Polymer electrolytics reduce ESR to 5mΩ-15mΩ while maintaining high capacitance, extending effective frequency to 500kHz." "The solution is multi-decade capacitor spacing: typical combinations are 10µF, 1µF, and 0.1µF in ratios of 10:1." "Modern FPGA designs targeting sub-5mΩ impedance use four decades: 100µF (VRM side), 10µF (regional placement), 1µF (IC side), 0.1µF (per power ball)." "Xilinx Ultrascale+ and Intel Stratix 10 specify per-rail impedance targets in their power delivery network design guides, typically 2mΩ-5mΩ for 0.85V-1.0V core rails switching 30A-50A."

  4. PCBSync, "PCB Power Integrity: PDN Design & Decoupling Strategy"
    What we used

    And "Power Bus Decoupling Strategy", "Power integrity is the discipline of keeping every chip's supply voltage stable while it slams current on and off, which on a real board means holding the power distribution network (PDN) impedance below a target value across a wide band of frequencies." "Power integrity comes down to one number and one shape: keep PDN impedance below the target impedance (Z_target = ΔV ÷ I_transient) and keep the profile flat, no resonant peaks." "A 1.0 V core with a 5% budget (50 mV) and a 10 A transient needs 5 mΩ." Target table: FPGA/SoC core 1.0 V ±5% 50 mV 10 A → 5 mΩ; high-current core 0.9 V ±3% 27 mV 25 A → ~1.1 mΩ; auxiliary/PLL 1.8 V ±5% 90 mV 3 A → 30 mΩ; I/O bank 3.3 V ±5% 165 mV 1 A → 165 mΩ. "Modern core rails run at 0.7–1.0 V with maybe 5% margin, so you have only tens of millivolts to play with, while a single FPGA can demand 50 A step changes in nanoseconds." "Above a few hundred MHz the board cannot help, that job belongs to package and on-die capacitance." "Second, target impedance only needs to hold up to the knee frequency of the current edge, f_knee ≈ 0.35 ÷ t_rise; there is little energy above it, so driving impedance down at 1 GHz is pointless." "A decoupling strategy is a target impedance held across a frequency band, not a pile of capacitors near the power pins." "To hold 5 mΩ at 100 MHz, your total loop inductance has to stay under 8 picohenries. A mounted 0402 with its vias is somewhere around 1.5–2 nH. That is 250 capacitors in parallel, which is why a large FPGA board genuinely carries two hundred decaps." "The impedance of an inductor is 2πfL. Solve it for the inductance that just reaches 5 mΩ: at 1 MHz you can afford 796 pH; at 10 MHz, 80 pH; at 100 MHz, 8 pH." "A 100 nF 0402 carries about 0.5 nH of ESL on its own before you mount it. Add pads, traces and the vias down to the planes and a realistic mounted loop is 1.5–2 nH." "At 2 nH per capacitor you need 250 in parallel to hold 5 mΩ at 100 MHz. At a well-mounted 0.65 nH you need 82." "The decade rule, 10 µF, 1 µF, 100 nF, 10 nF, puts an impedance peak between every adjacent pair." "f_anti = 1 / (2π√(L_big × C_small)). Run it on the classic pairing, a 10 µF 0805 at 0.9 nH beside a 100 nF 0402, and the peak lands at 16.8 MHz." "A 10 µF 0805 at 0.9 nH with a 100 nF 0402 at 0.5 nH puts the anti-resonance at 16.8 MHz while the 100 nF's SRF is 22.5 MHz. At 16.8 MHz the 100 nF is still firmly capacitive, the tank is real, and so is the peak."

  5. hilelectronic, "PCB Power Distribution Network Design"
    What we used

    "The PDN bandwidth requirement is set by the fastest current transient the load can produce: fmax = 0.5 / trise. For a processor with 1 ns current rise time: fmax = 500 MHz. For DDR5 at 6400 MT/s with 100 ps edge rates: fmax extends to 5 GHz." "The design goal is to maintain Z < Ztarget continuously from DC through fmax. Any frequency gap where PDN impedance exceeds Ztarget is a potential source of voltage droop, oscillation, or logic failure at the corresponding switching frequency." Target table: DDR5 DRAM VDD 1.1 V 3% 5 A → 6.6 mΩ; FPGA core 0.9 V 3% 30 A → 0.9 mΩ; PCIe endpoint 3.3 V 5% 3 A → 55 mΩ; high-speed SerDes 1.8 V 2% 2 A → 18 mΩ; GPU core 1.0 V 2% 200 A → 0.1 mΩ

  6. allpcb, "Mastering PDN Impedance"
    What we used

    "If a resonance peak appears at 50MHz with an impedance of 100mΩ, far above the 5mΩ target, it indicates a need for design adjustments, such as adding decoupling capacitors."

  7. Research-DAQ, "Power-tree and PDN design"
    What we used

    "Ztarget = DeltaV / DeltaI. Example: a 1.0 V FPGA rail may allow 20 mV transient deviation for a 1 A step: Ztarget = 0.020 / 1 = 20 milliohm." "A capacitor is not ideal: Zc = ESR + j*omega*ESL + 1/(j*omega*C)." "Multiple capacitor values can create anti-resonance peaks rather than a universally lower impedance." "The current loop area from device pin through capacitor and ground return is more important than the visual distance alone."

  8. Demo arithmetic verified in Python before writing: 3.0 V × 5% = 150 mV budget; Ztarget = 150 mV / 0.1 A = 1.5
    What we used

    Ω; twin sag 0.36 V is 2.4× the budget; rail 3.0 − 0.36 = 2.64 V, 0.06 V below the 2.7 V sensor minimum; predicted band 0.30–0.40 V is 2.0×–2.67× the budget; measured Z = 0.36 / 0.1 = 3.6 Ω, 2.4× over target; SRF 100 nF 0402 at 0.5 nH = 22.5 MHz (reproduces AtlasPCB), at 450 pH = 23.7 MHz (AtlasPCB's table says 24 MHz); 100 nF 0805 at 1.2 nH = 14.5 MHz (Flux says 14 MHz); 100 nF 0402 at 0.7 nH = 19.0 MHz (Flux says 19 MHz); 10 µF 0805 at 0.9 nH SRF = 1.68 MHz; 1 µF 0402 at 450 pH = 7.5 MHz (AtlasPCB says 7.5 MHz); 10 nF 0201 at 250 pH = 100.7 MHz (AtlasPCB says 100 MHz); X at 48 MHz: 0.5 nH = 0.15 Ω, 8 nH = 2.41 Ω, 2 nH = 0.60 Ω, 1 nH = 0.30 Ω; X at 200 MHz, 1 nH = 1.26 Ω (reproduces AtlasPCB); X at 100 MHz, 8 pH = 5 mΩ (reproduces PCBSync); hub plane capacitance 100 × 60 mm, 0.1 mm, εr 4.2 = 2.23 nF (AtlasPCB's 100 × 80 mm example is 3.0 nF); X of 2.23 nF at 48 MHz = 1.49 Ω, at 200 MHz = 0.36 Ω, at 1 GHz = 0.07 Ω; anti-resonance 10 µF 0.9 nH + 100 nF = 16.8 MHz (reproduces PCBSync); f_knee for 1 ns = 350 MHz, fmax = 500 MHz; Andwin 3.3 V 5% 8 A = 20.6 mΩ, 1.0 V 5% 10 A = 5 mΩ; Flux 1.2 V 5% 2 A = 30 mΩ; AtlasPCB 0.9 V 3% 30 A = 0.9 mΩ; buck D = 5/24 = 0.2083, ΔIL = 0.792 A (reproduces #44); droop budget 10 + 10 + 10 + 15 + 5 = 50 mV (reproduces AtlasPCB)

More in precision measurement

Sensor Calibration and DriftA gyroscope measures rate, not angle, the angle is the integral of the rate, so any bias is integrated too: a bias of 0.1°/s adds 6° of error every minute, 360° in an hour. Bias is the offset the sensor reads at rest; it shifts with temperature and time, and random noise adds its own slow wander (a random walk whose standard deviation grows with the square root of time). The fixes are simple and in order: measure the bias at startup while the device is still and subtract it; recalibrate continuously when the device rests; fuse with an accelerometer (complementary filter, α ≈ 0.98), the accelerometer gives tilt with no drift when still, the gyro gives smooth fast motion. Yaw has no gravity reference, so it needs a magnetometer. Even a well-calibrated gyroscope cannot be trusted for hours without a second sensor to check against.24 minThe Spec Sheet That Lies About AccuracyA sensor can be precise without being accurate, a tight cluster of readings, far from the true value. Accuracy is closeness to the true value; precision is how tightly repeated readings group; resolution is the smallest detectable change. The Levelogger 6 M100's resolution is 0.6 mm on a 100 m full scale, 83 times finer than its ±5 cm accuracy. Averaging fixes random noise; it cannot fix systematic error. That needs calibration against a known reference.22 minThe Ground That Isn't a GroundOn the schematic every ground symbol is the same point. On the board they are joined by copper, and copper has resistance and inductance, so the ground at the sensor is not the ground at the ADC. The fix is a plane, a return under the trace, and one bridge between the noisy and the quiet.24 minEMC pre-compliance testingAbout half of products fail EMC testing on the first attempt, and each failure costs $5,000–$30,000 and 2–6 weeks. A pre-compliance scan on your own bench finds the failure before the lab does, for $300–$800 a scan.30 minThe Rail That Doesn't RingYour 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.20 min