SIPI

Power Integrity / 05

PDN Frequency Regions: Board, Package, and Die

Apply a load step and the rail does not sag once. It sags three times, on three very different timescales, because three different stores of charge run out one after another before the next takes over. Knowing which droop you are looking at tells you which hardware can possibly fix it — and, more usefully, which cannot.

One load step · three droops, three owners
nominal Vdd lower limit first droop on-die C second droop package + board caps third droop bulk caps, then the VRM 1 ns 10 ns 1 µs 100 µs time (log) V load step
The rail does not sag once. Apply a current step and it dips three times on three timescales, because each stage of the network runs out of stored charge before the next takes over. Each dip has a different owner and a different fix, and they barely overlap: the first droop at around a nanosecond is held up by on-die capacitance and can only be fixed on the die or in the package; the second at tens of nanoseconds belongs to package capacitors and board ceramics; the third at microseconds is bulk capacitance and then the regulator's control loop. This is why identifying which droop you have is the first move. A first-droop problem cannot be fixed with a board capacitor no matter how many you add or how close they sit — the inductance between the board and the die is a fence, and the charge simply cannot get there in time.

Why one step produces three dips

When you capture a high-speed load transient on an oscilloscope, you do not see a single, smooth voltage sag. Instead, you see a jagged sequence of three distinct dips unfolding across vastly different timescales: one in the first nanosecond, another tens of nanoseconds later, and a third stretching across microseconds. To an untrained eye, this looks like random ringing or chaotic noise. But it is actually an orderly organizational chart of the system's power architecture.

Power integrity is fundamentally a division of labor: a strict jurisdictional map where the VRM, board capacitors, package planes, and on-die capacitors each “own” a specific frequency band. Because each tier is separated from the transistors by a different amount of physical distance and loop inductance, each tier runs out of stored charge before the next slower, larger tier can respond. Knowing which hardware tier governs each timescale is the most critical diagnostic rule in PI: an engineer who tries to solve a 1 ns droop with a board capacitor will spend weeks achieving nothing, because that event sits entirely outside the board's jurisdiction.

Picture the instant a load step arrives. The current has to come from somewhere, and the candidates are at different distances — on-die capacitance a few micrometres away, package capacitance a millimetre away, board capacitors centimetres away, the regulator further still.

Between each of those and the load is some inductance, and inductance sets how quickly that source can start delivering. So the nearest store responds first, because it has the least inductance in the way. It is also the smallest, so it runs down quickest — and the rail sags until the next store, further away and slower, picks up.

That handover is a droop. There are three of them because there are three handovers, and each one happens on the timescale set by the inductance between those two stores. The pattern is a relay race where each runner is faster than the next but carries less.

Three droops, three owners

A useful decomposition, not a law of nature The three bands below describe a common architecture — a board regulator, a package with its own capacitance, a die with on-die decoupling — and the numbers move with every one of those choices. Nothing forces there to be exactly three. A design with no package capacitance merges two of them; one with an on-package regulator loses the longest; a distributed plane contributes a spread of overlapping resonances rather than a single clean peak, and peaks can appear within a stage rather than only at the boundaries between them.

What does generalise is the reasoning: charge for a fast transient must come from somewhere electrically close, and series inductance decides how close that has to be. Use the table as a way to ask which store is supplying your event, not as a set of fixed frequencies.

Apply a load step and watch the rail. It does not sag once; it sags three times, on three timescales, and each dip is a different part of the network running out of stored charge before the next one takes over.

DroopTimescaleFrequencyHeld up byFixed by
First~1 nshundreds of MHzon-die capacitancedie and package only
Second~10–100 nstens of MHzpackage caps, board ceramicspackage, board mounting
Third~1–10 µs~100 kHzbulk caps, then the VRMbulk capacitance, VRM loop

Identify which droop you are looking at and the list of things worth trying collapses to a handful. Skip that step and you can spend weeks adding capacitors on the wrong side of a fence.

The fence is inductance

Every stage is separated from the next by series inductance — balls, vias, bond wires or bumps, plane spreading. That inductance rises with frequency as an impedance, and above some frequency it is simply larger than anything the far side can offer. Charge on the board side physically cannot get to the die fast enough, and the die is on its own.

The rule this produces You cannot fix a die-frequency problem with a board component. Not with a better capacitor, not with more of them, not with perfect placement. If the first droop is out of budget the answers are: more on-die capacitance, more or better-distributed power bumps, lower package inductance, or reduce the di/dt at source with clock gating and staged power-up. That is the whole list.

The converse is just as useful and much cheaper to act on. If the problem is a third droop — microseconds, a workload transition, a big power-state change — then on-die capacitance is irrelevant to it and the die team has nothing to contribute. It is bulk capacitance and VRM response, and it is a board and regulator conversation.

Where this gets genuinely hard is that the handoffs between stages are exactly where the impedance peaks are, so a problem can sit in two owners' territory at once. That is the case chip-package-system co-analysis exists for: neither side's model, simulated alone, contains the resonance.

Why this decides how your board behaves

What to do about it

Measure the timescale first, always. Everything else follows from it. A nanosecond dip and a microsecond dip look similar on a badly scaled plot and have nothing in common as problems.

For the first droop, the only levers are on-die and in-package. More on-die decoupling capacitance, lower package inductance, or reducing the di/dt itself — staggering clock-gate release so the step is spread over more cycles is often the cheapest fix available and costs no hardware at all.

For the second, attack mounting inductance. This is the band where board ceramics and package capacitance hand over, so the loop inductance of both is what sets it.

For the third, look at the regulator's loop. Its bandwidth decides where it stops being able to help, and bulk capacitance covers the gap between that and the ceramics. See VRM loop bandwidth.

Using this to find a fault
  • Read the dip's width, not its depth, to identify it. Width names the owner; depth only tells you how bad it is.
  • A droop that did not change when you added board capacitors. It was the first droop. Stop adding parts.
  • Ringing on the recovery rather than a clean settle. There is an anti-resonance between the two stages handing over. Its frequency identifies which two.
  • Failures that only happen on a workload transition. Look at power-state exits and clock-gate releases — the largest di/dt on a modern SoC is almost never the steady workload.
Go deeper — what sets each timescale, and why this decomposition is a convenience

Each handover happens at roughly the resonant frequency of the inductance between two stores and the capacitance of the nearer one:

f ≈ 1 / ( 2π √( Lbetween · Cnearer ) ) the first droop is package inductance against on-die capacitance; the second is board and mounting inductance against package capacitance

That is the same expression as the anti-resonance frequency, and deliberately so: a droop in time and a peak in frequency are the same event seen two ways. If you have an impedance curve and a transient plot of the same PDN, each droop corresponds to a peak, and their frequencies match. Checking that they do is a good way to validate a model.

The decomposition is a convenience, not a law. The three bands describe a common architecture — a board regulator, a package with its own capacitance, a die with on-die decoupling — and the numbers move with every one of those choices. A design with no package capacitance merges two of them. One with an integrated voltage regulator loses the longest. A distributed plane contributes a spread of overlapping resonances rather than a single clean peak, and peaks can appear within a stage rather than only at the boundaries between them.

What does generalise is the reasoning: charge for a fast transient must come from somewhere electrically close, and series inductance decides how close that has to be. Use the bands as a way to ask which store is supplying your event, not as a set of fixed frequencies.

One practical consequence of the time-frequency equivalence: a transient measurement is often easier than an impedance measurement and tells you much of the same thing. A step response with a clean scope shot gives you the droop timescales directly, and from those you can infer where the peaks are without ever connecting a VNA.

In the real world

The three-droop picture is the most useful diagnostic frame in power integrity, and its value is almost entirely in ruling things out. Most PI debugging time is lost to fixing the wrong band — adding board capacitance to a package problem, or re-tuning a regulator for a nanosecond event.

So the first question on any rail problem is not “how big is it?” but “how long is it?” One measurement, and two thirds of the possible fixes are eliminated before you have opened the layout tool.

Related