← Geometric Power Topologies for FPGAs
Chapter 5 · Living chapter
Control Boards and Falsifiable Measurement Discipline
Book 1: Geometric Power Topologies for FPGAs — where simulation ends and physics begins.
Chapter 5 · Living chapter
Book 1: Geometric Power Topologies for FPGAs — where simulation ends and physics begins.
Status — pre-fabrication
Falsifiable — This chapter defines hypotheses and acceptance criteria that can be disproven by measurement.
Simulation is a claim about a model. Fabrication is a question asked of reality. Chapter 3 showed what the toroidal PDN does inside a lumped model; Chapter 4 showed the Hilbert layer carries at most 0.41% of return current and must therefore do whatever it does in the field domain, not the current domain. Both chapters end at the same wall: no amount of re-running the netlist tells you whether the geometry matters on a physical board. This chapter is about how we find out — and, more importantly, about the discipline that makes the answer mean something.
The cheapest way to fake a result in hardware is to build only the thing you believe in. A board with a toroidal VCC ring and a Hilbert ground, measured alone, will produce numbers. They will be meaningless. Without a control, "the EMI is 6 dB lower" has no referent: lower than what? A board from a different fab lot? A different stackup? A Tuesday?
The rule this project runs on: a claim without a control is an anecdote. So every test vehicle ships in a pair. The control board is identical to the experimental board in every variable we are not testing — same 100×100 mm outline, same four-layer stackup, same copper weights, the same twelve 100 nF decoupling capacitors at the same coordinates, the same stitching-via grid, the same drill, the same non-electrical FPGA placeholder, the same signal-source placement. The only differences are the two artworks under test: the L2 ground (Hilbert order-4 curve versus solid plane) and the top-layer power distribution (toroidal ring versus solid pour).
And one more variable that hides more frauds than any other: both boards are ordered in the same fab lot. Process variation between lots — copper thickness, dielectric tolerance, etch profile — can easily exceed the effect being measured. Same-lot pairing is the only way the comparison isolates the artwork.
Before any copper is ordered, the hypotheses are written down in a form where they can fail:
Note what is missing: there is no hypothesis that says "the geometry helps." Each hypothesis is a question with a permitted answer of no. And the null result is explicitly a valid result: "no measurable benefit" resolves the claim as thoroughly as a confirmation would. A research program that cannot be disappointed is not research; it is merchandising.
| Measurement | Equipment | Method |
|---|---|---|
| PDN impedance | 2-port VNA, ≥10 kHz–1 GHz | 2-port shunt-through, Z = 25·S21, with a common-mode choke (Picotest J2102A-class) in the receiver loop to kill the ground-loop artifact |
| Near-field emissions | Spectrum analyzer, 30 MHz–1 GHz, H-field loop and E-field stub probes | Fixed-height 5 mm grid scan, 5 mm pitch, fixed probe orientation |
| Radiated (pre-compliance) | TEM/GTEM cell, 150 kHz–1 GHz | Identical fixture position and orientation; report deltas in dB |
Three honesty notes the table hides. First, the 2-port shunt-through method is chosen because a single-port VNA measurement of milliohm impedances is dominated by fixture artifacts; the shunt-through with common-mode suppression is the industry-standard way to see below 100 mΩ without lying to yourself. Second, the budget alternative — a nanoVNA-class unit — can profile impedance coarsely, but its dynamic range below ~10 mΩ is poor, and the chapter says so rather than pretending the cheap instrument equals the good one. Third, the stimulus is a canned crystal oscillator on a battery or linear supply, deliberately not a switching regulator: we are testing geometry, not power-supply noise, and everything not under test must be boring.
The acceptance criteria are written down before the boards exist — pre-registration, the same discipline clinical trials use, because it removes the temptation to decide what counts as success after seeing the data.
| # | Metric | PASS if… | Meaning |
|---|---|---|---|
| A1 | PDN impedance vs control | Experimental |Z| ≤ 1.5× control across 1–200 MHz | The Hilbert layer does not catastrophically degrade the PDN |
| A2 | Near-field maximum | ≥6 dB reduction in a defined band at least one decade wide, to claim benefit; ≤3 dB change = "no effect" | Resolves H1 |
| A3 | Near-field average | Report band-by-band deltas with measurement uncertainty (±2 dB typical) | Resolution, not judgement |
| A4 | PDN resonance shifts | Document peak frequency and delta vs control | Explains mechanism (inductance added → H2) |
| A5 | Toroidal vs solid top | Ripple-band ΔZ within ±20% at the capacitor ring | Resolves H3 |
| A6 | Repeatability | Repeat-mount spread ≤2 dB; otherwise investigate the fixture before concluding anything | Data-quality gate |
Read A2 carefully, because it is the discipline in one row. Six dB in a decade-wide band is the minimum to claim a benefit at all. Three dB or less is not a win with smaller numbers — it is no effect, because ±2 dB is inside typical measurement uncertainty. The table decides in advance that most "improvements" people report are noise, and refuses to let us become one of them.
Falsifiability is not a mood; it is a list of outcomes, written before the experiment, that would make us abandon each claim:
And one rule for language: claims like "700% EMI cancellation" remain marketing until A1–A6 say otherwise. If a number cannot be traced to a measurement in the results document, it does not appear in this book's future editions — that standard is what Chapter 1 promised and this chapter enforces.
The measurement campaign is not a fresh start; it is the second half of an argument the simulations began:
That is the point of the whole discipline: simulation narrows the search space and tells you where not to look; the controlled measurement tells you whether the thing that remains actually exists.
The boards are designed, generated, and frozen — full Gerber sets for experimental and control, validated with zero errors, sitting in the project bundle (they do not expire). The measurement plan above is versioned and agreed before fabrication, exactly as this chapter documents. No boards have been fabricated and no physical measurements have been taken. The fab budget is the gate, not the physics: when the pair is ordered, they go in the same lot, two of each for build variation, with the experimental L2's sparse copper flagged to the fab's DFM review as a documented caveat.
Nothing in this chapter is ESTABLISHED as a physical fact yet, and the chapter's tables are the machinery for finding out. What is established: the controls, the criteria, and the refusal to publish a geometry claim that cannot survive them.
Chapter 5 — draft of 2026-10-02. The measurement plan, acceptance criteria A1–A6, and test-vehicle specifications are drawn from the project's pre-fabrication measurement plan (version 1.0, defined before fabrication). Test vehicles: 100×100 mm, 4-layer, toroidal VCC top + order-4 Hilbert L2 versus solid-plane control, twelve 100 nF 0402 caps, same-lot fabrication requirement. All claims in this chapter are design-stage; no physical measurements exist. The hypotheses (H1–H3) and acceptance criteria are reproducible from the downloadable test-vehicle bundle.