SI/PI Measurement: Fixtures, Calibration, and Instrument Effects

When an engineer connects a coaxial cable to a vector network analyzer (VNA) or clips an oscilloscope probe onto a test point, there is an instinctive tendency to trust the resulting waveform as absolute truth. If the oscilloscope shows a 2 GHz ringing oscillation, the natural assumption is that the transmission line is reflecting. If the VNA displays 4 dB of loss at 10 GHz, it is assumed the board trace attenuated the signal.

In high-speed hardware, however, an instrument never measures your circuit board in isolation. It measures your circuit board in series and in parallel with the measurement apparatus itself. The probe tip adds parasitic capacitance that slows down the rising edge; the three-inch alligator ground wire forms a lumped inductor that resonates violently with that capacitance, generating false ringing that does not exist on the board. The coaxial cables connecting the VNA introduce insertion loss, phase shift, and connector reflections that dwarf the device under test.

High-speed measurement practice is the art and discipline of removing the instrument from the data. It begins with rigorous calibration standards that mathematically subtract cables and fixtures to place the reference plane exactly at the device pins, and extends to specialized topologies — such as two-port shunt-through impedance sensing with common-mode isolation — that make it possible to measure sub-milliohm power distribution networks without being blinded by ground loops.

The measurement is part of the circuit

A scope probe participates in the circuit. Its input capacitance loads the node and its ground path adds inductance; on a fast or high-impedance node, that loading can materially change the waveform being measured.

Read the displayed value together with a second question: "what did the instrument and connection contribute, and how large is that effect relative to the behavior under investigation?"

Probe loading

Two parasitics matter, and the second is the one that surprises people.

Tip capacitance loads the node and slows the edge you were trying to measure. A few picofarads on a fast net is a real discontinuity.
Ground lead inductance is worse. The classic alligator-clip ground is tens of nanohenries, and it forms a resonant circuit with the tip capacitance — a resonance that lands squarely in the band you care about. When you see ringing, that is the first thing to rule out, and shortening the ground path is the test: if the frequency moves, it was yours. But it is not the only cause, and deciding it was without checking is how a real problem gets dismissed. The board can genuinely ring from a reflection, the probe can load a high-impedance node enough to create a resonance that is neither purely its own nor purely the board's, an under-damped supply rail can ring at the probe tip because it is ringing everywhere, and an over-driven front end can produce something that looks like ringing and is clipping. The discriminating checks are cheap: change the ground length, move the probe to a second point, halve the vertical scale, and compare against a second probe of a different type.

Two quick loading checks Move the probe and see whether the waveform changes. Probe a known-quiet node and see whether it is quiet. A change does not identify the whole cause, but it shows that the probe, connection, or reference path is part of the result.

TDR: resolution is set by the edge

A TDR launches a step and watches what comes back. Its ability to separate two features is set by how fast that step is:

resolution ≈ t_r · v / 2 the /2 is the round trip — and v depends on your assumed D_k, so every distance inherits that assumption

A slow TDR merges features a fast one separates, and reports their combined impedance as if it were one structure. So a "clean" TDR trace from a 50 ps instrument may simply be one that cannot see the problem.

Report the measured delay and, when distance is useful, the converted length together with the propagation velocity or dielectric assumption. This keeps the instrument observation separate from the material-based conversion.

VNA: calibration decides where the measurement starts

Calibrating moves the reference plane — the place the instrument believes its measurement begins. Everything between that plane and your device is included in the answer.

SOLT uses known short, open, load and thru standards. Convenient, and only as good as the standards' models — which degrade at high frequency.
TRL uses thru, reflect and line standards fabricated on the same board as the device, so they share its manufacturing. It shifts the accuracy question from "how good is the standards model" to "how well did this coupon come out", and that is usually the better trade at high frequency — but not always, and not by default. TRL needs the line standards to be uniform and their lengths known, it has band gaps where a line is a multiple of a half wavelength, and a coupon that etched differently from the device carries that difference straight into the result. A well-characterised electronic calibration unit can beat a poor coupon comfortably. Costs coupon area, and needs several line lengths to cover a broad band.

Then there is de-embedding, which moves the plane further — onto the device itself. That is a separate operation with its own error, and its signature is ripple rather than offset. See de-embedding.

Measuring a PDN: why a one-port setup loses sensitivity

A PDN's impedance is milliohms. A one-port VNA measurement of a milliohm load is dominated by the cable and connector resistance, which is also milliohms — so the instrument is mostly measuring itself.

A common solution is two-port shunt-through: drive through one port, sense with the other, both connected across the same point, and compute the impedance from S₂₁ rather than S₁₁. The sense path carries much less current than the drive path in the low-impedance limit, which greatly reduces its series-resistance contribution.

Z ≈ ( Z₀ / 2 ) · S₂₁ / (1 − S₂₁) valid while Z is far below Z₀/2, which is exactly the regime a PDN lives in

One practical catch: with both ports grounded at the same place, a ground loop through the instrument can dominate the reading at low frequency. A common-mode choke on one cable is the usual fix, and its absence is the usual reason a PDN measurement looks wrong below a megahertz.

BERT: what a run of a given length can support

A bit error rate tester directly observes error events under its configured pattern, receiver, and sampling conditions. Its limit is arithmetic: observing no errors in N bits supports a bound of about 3/N, and nothing stronger.

At 16 GT/s, a 95% bound near 10⁻¹² takes about three minutes of error-free traffic at one sampling position. A bathtub has dozens of positions and sign-off wants every corner, which is how three minutes becomes days. See bathtub curves.

Repeatability comes before accuracy

Before comparing a measurement to anything, find out how much it varies when nothing has changed. Three levels, cheapest first:

Same fixture, reconnected. Catches connector repeatability, which at high frequency is often the largest term.
Different fixture, same board. Catches fixture-to-fixture variation and de-embedding error.
Different board, same design. Catches manufacturing spread — which is a property of the design, not the measurement.

That spread is a practical repeatability floor: differences below it cannot be distinguished confidently without more repetitions or a better uncertainty model. Establishing it prevents ordinary setup variation from being assigned to the simulation or hardware.

Using this to find a fault

Ringing that changes when you move the probe. The probe's ground lead. Use a short spring tip before believing anything.
A PDN impedance that looks flat and too high at low frequency. Ground loop through the instrument. Add the choke.
A TDR trace with no features where you expected one. Check the instrument's rise time against the feature's size before concluding it is not there.
A de-embedded S₂₁ above 0 dB. A passive fixture-plus-device cannot have gain, so something in the de-embedding is wrong — but "removed it twice" is only the most common of several explanations, and they are distinguishable. Over-de-embedding gives an excess that grows with frequency, because it is the fixture's own loss applied twice. A fixture model whose impedance is wrong gives a ripple that oscillates with frequency at the fixture's electrical length. A calibration plane in the wrong place gives an excess that tracks the phase error. And a noise floor problem gives something that is above 0 dB only where the raw measurement was already near the floor. Plot the excess against frequency before concluding which one you have.
Results that differ between two runs. Establish repeatability before anything else. It is the cheapest measurement on this page and it decides whether the rest are worth doing.

Go deeper — why shunt-through works, and what each instrument cannot see

The shunt-through trick, and what port 2 actually is. In a one-port measurement the cable's series resistance sits in the same path as the current through the device, so the two are indistinguishable. Two-port shunt-through separates them: port 1 drives current through the device and port 2 measures across it, so the resistance of port 1's cable is no longer in the sensing path.

Port 2, though, is terminated in 50 Ω like any VNA port — it is not a high-impedance voltmeter and it does carry current. The reason the method works on a low impedance is a ratio, not an open circuit: when the device under test is a few milliohms and the port is 50 Ω, the current into port 2 is four orders of magnitude below the current through the device, so port 2's cable resistance contributes negligibly to the voltage it reports. Raise the device impedance towards 50 Ω and that stops being true, which is why shunt-through is a low-impedance technique rather than a general one. The exact relation for the ideal shunt topology with equal real reference impedances is:

Z = Z₀ · S₂₁ ⁄ [2 · (1 − S₂₁)] and for |Z| ≪ Z₀ this simplifies to Z ≈ Z₀ · S₂₁ ⁄ 2, which is the form usually quoted — but it is the approximation, not the definition

How far the approximation can be trusted is a question with a number attached, rather than a matter of intuition — and the useful range is relatively narrow. To first order the simplified form overstates |Z| by 2|Z| ⁄ Z₀, so on a 50 Ω instrument the error reaches 1% at 0.253 Ω and 10% at 2.78 Ω. At 1 Ω it is already 3.85% wrong. Below a few hundred milliohms — the range a PDN actually lives in — the two agree to 0.4% or better and the choice genuinely does not matter. Above about a quarter of an ohm, use the full expression; it costs one division.

Two further things the simplification hides. The exact relation is complex: S₂₁ has phase, and an impedance that is part inductive — which every real capacitor mounting is above its self-resonance — needs the complex division to recover both the magnitude and the sign of the reactance. Taking |S₂₁| and scaling it gives a magnitude and throws the reactance away, which is exactly the information you were measuring for. And the relation assumes the ideal shunt topology with equal real reference impedances at both ports; a fixture with a series element, or ports at different reference impedances, needs its own algebra.

It is a genuine four-terminal arrangement in the sense that matters — the drive and sense paths are separate — but calling it a Kelvin measurement invites the wrong mental model, because a Kelvin voltmeter draws no current by construction and this one draws little by circumstance. A high-impedance receiver configuration is a different setup with different error terms, and the two should not be conflated. And none of this removes the need to think about the ground return between the two ports: the semi-floating differential amplifiers and common-mode chokes that appear in low-impedance fixtures exist because the port-to-port ground loop is the dominant error at low frequency, not the cable resistance.

What each instrument structurally cannot see.

A VNA measures small-signal, steady-state, linear behaviour. It cannot see non-linearity, and it cannot see anything that only happens while the circuit is switching.
A TDR is most sensitive to impedance changes. Distributed loss changes amplitude and edge shape but is difficult to localise as one feature. A large early discontinuity can also mask smaller structures behind it.
A scope sees a voltage against its own ground reference. If that reference is moving — ground bounce — the scope cannot tell that from a signal change.
A BERT counts errors under a defined stimulus and receiver configuration. It establishes performance for that run but needs other measurements to identify the mechanism.

The redundancy between them is what makes correlation possible. A TDR trace is related to the step response of S₁₁, so a TDR and a VNA can be transformed into comparable views when their bandwidth, reference plane, impedance, windowing, and fixture treatment match. A disagreement first prompts a check of those conditions.

VNA calibration standards: SOLT vs. TRL vs. E-Cal. The quality of any VNA measurement is strictly limited by its calibration method, which defines where the mathematical reference plane sits:

SOLT (Short, Open, Load, Thru): The classical coaxial calibration standard. Relies on a known precision 50 Ω termination resistor, an open circuit with calibrated fringing capacitance coefficients, and a short circuit with known residual inductance. Best for: Coaxial test fixtures with SMA, 2.92 mm, or 1.85 mm precision connectors. Limitation: Fails on printed circuit boards and wafer probes because fabricating an ideal, pure 50 Ω surface-mount resistor standard that remains parasitic-free up to 40 GHz is physically impossible.
TRL (Thru, Reflect, Line): The gold standard for planar PCB and substrate fixtures. Instead of requiring an impossible 50 Ω load, TRL uses a non-zero length transmission line (“Line”) fabricated on the exact same board layer as the device under test. The characteristic impedance of this line becomes the VNA's reference impedance, and the phase difference between the Thru and Line establishes the frequency scale. Best for: Moving the calibration reference plane directly onto the PCB surface, directly at the pins of a chip or connector launch. Limitation: Each Line standard is only valid over an 8:1 frequency span (where phase delay is between 20° and 160°); covering 100 MHz to 40 GHz requires fabricating multiple lines of different lengths.
E-Cal (Electronic Calibration): A USB-controlled solid-state calibration module containing PIN diodes or FET switches that cycle through precision internal impedance states. Best for: Speed, repeatability, and bench productivity. A full 4-port calibration completes in seconds with a single connection, eliminating operator torque-wrench errors and physical standard wear.

Breaking the PDN ground loop: why common-mode chokes are mandatory. In two-port shunt-through measurements below 1 MHz, engineers frequently observe that the measured impedance flattens out into a completely false resistive floor of 10 to 50 mΩ, hiding true milliohm PDN behavior.

This error is not caused by the VNA's internal noise floor. It is caused by a ground loop formed by the braided outer shields of the two coaxial cables and the VNA's chassis ground. Because the shield resistance of a typical 1-metre RG-142 cable is around 10 to 20 mΩ, return current from Port 1 splits between the device under test and the outer shield of Port 2.

To eliminate this floor and accurately measure sub-milliohm power planes down to 10 kHz, you must insert a common-mode coaxial choke (or a high-bandwidth RF isolation transformer) in series with the sensing cable on Port 2. By wrapping the coaxial cable around a high-permeability toroidal ferrite core, the choke adds several microhenries of common-mode inductance to the outer shield without affecting the differential TEM signal inside the cable. This forces all return current to flow through the device under test, lowering the measurement floor down to 100 μΩ.

In the real world

The most expensive measurement errors are not wrong numbers; they are right numbers about the wrong thing. A perfectly calibrated VNA measuring to an undocumented plane can produce a file that is difficult to compare with simulation because the two reference planes may differ.

Package the data with what was measured, the reference plane, calibration and de-embedding, instrument settings and bandwidth, fixture and probe connection, environmental conditions, and observed repeatability. That compact record lets another engineer reproduce the setup and decide which comparisons are valid.

Correlating Simulation and Measurement — what a disagreement can mean
S-Parameter De-Embedding and Fixture Removal — moving the reference plane
S-Parameters: Insertion Loss, Return Loss, and Mixed Mode — what the VNA produces
Target impedance — the milliohms shunt-through exists to measure
Units and conventions — the three questions to ask of any number
Lab D: ADC Interference, Aliasing and Reference Noise — what a noisy or offset reading means when the instrument is doing the sampling