How many decoupling caps?
How do you decide how many decoupling capacitors a processor needs, and how do you verify the PDN?
| Level | ●● onsite technical |
| Track | Power |
| Really testing | Target impedance instead of ritual. |
Free full answer — one of 12 questions answered in full on this site.
Show the answer
Say this first: From the target impedance, not from "one 100 nF per pin". Z_target = V·ripple% / ΔI. Then pick capacitor values and counts so the PDN impedance stays below target from the VRM's bandwidth up to the frequency where the package and die capacitance take over — then verify by simulation and measurement.
Then explain
- 0.85 V, 3%, 20 A step → ~1.3 mΩ.
- Each MLCC is only effective near its SRF; its mounting inductance (vias, pads: ~0.5–1 nH) dominates above. Use several values to cover the band, avoid sharp anti-resonance peaks, use low-inductance mounting (vias in or beside pads, short connection to the planes).
- Verify: post-layout PDN simulation (Z vs f); bench: 2-port VNA shunt-through measurement for mΩ levels; load-step test with a fast transient load.
Going deeper
Why counting capacitors fails — worked. Target 1.3 mΩ. One 22 µF 0805 mounted with about 0.6 nH has an impedance of 2π × 10 MHz × 0.6 nH ≈ 38 mΩ at 10 MHz. Reaching 1.3 mΩ at 10 MHz would take about 30 of them in parallel. That is why the job is split by frequency:
- below ~100 kHz — the regulator's loop;
- ~100 kHz to ~1 MHz — bulk and polymer caps;
- ~1 MHz to tens of MHz — MLCCs, limited by their mounting inductance;
- above ~100 MHz — package and on-die capacitance. Board caps can't help there.
Anti-resonance. Two cap values in parallel make a parallel LC between their SRFs — a peak. Use more values close together, or caps with some ESR, to damp it.
Where people go wrong
- "One 100 nF per pin" with no target.
- Connecting caps with traces instead of vias straight into the planes — the trace inductance wins.
- Simulating the caps but not the VRM's real output impedance.
More like this in the book
This is PI-01 from Hardware Engineer Interview Questions. The book has 5 questions on power integrity & protection and 141 in all, each answered the same way: what to say first, how to explain it, the numbers worth quoting and the follow-ups that come next.
Errata
Correction: Appendix A target-impedance example corrected from ~1 mΩ to ~1.3 mΩ, to match this question. Fixed before the first printing. All errata
Ask about this question
Stuck on a step, or think something is wrong? Send it with PI-01 in the subject. Good questions get answered on this page, and corrections go into the next printing.