Measurement: Multimeter, Oscilloscope, and Logic Analyzer
Every instrument is part of the circuit while it measures it. The question is how much.
🎯 By the end you'll understand
Why a multimeter's 10 MΩ input reads a high-impedance divider low, and why the same meter switched to a current range inserts a resistor that can brown out the board it is measuring. Why 8 µA and 40 mA in the same capture is a five-thousand-to-one dynamic range problem that no single shunt resistor solves, and what people use instead. Oscilloscope bandwidth as a rise-time filter, sample rate and memory depth as separate constraints, and probe capacitance as a component you soldered onto the node. Why a long probe ground lead rings, at what frequency, and why shortening it makes the ringing disappear rather than fixing anything. When a logic analyzer answers the question and when only a scope can. And exactly how Beacon's canonical 8 µA and 45 ms at 40 mA were obtained.
KEY INSIGHTThe instrument is a component you added to the circuit, and its value is on its datasheet
The ringing was beautiful, repeatable, and about thirty megahertz. Two engineers spent a morning designing a snubber for it. Then someone swapped the probe's ground clip for a short spring and the ringing was gone — because it had never been in the board.
Measurement in embedded work is not a neutral act of observation. A voltmeter is a resistor you connected in parallel. An ammeter is a resistor you connected in series. A scope probe is a capacitor you soldered to the node, hanging off an inductor you formed with the ground lead. Each of those is a real component with a real value, and each one changes the circuit it is measuring — sometimes negligibly, sometimes decisively, and the difference between those two cases is arithmetic you can do before you touch anything. The engineers in the story above were not careless; they were reading a real waveform on a real screen. The waveform was of the probe, not of the board, and the only defence is knowing what your instruments are made of.
Four instrument parameters cause most wrong conclusions in embedded debugging. Meter input impedance (typically 10 MΩ) loads high-impedance nodes — Thevenin's theorem from Lesson 1.3 tells you exactly how much. Burden voltage is the drop across the shunt inside an ammeter, and on microamp ranges that shunt can be kilohms, which is enough to stop a board from starting. Bandwidth limits how fast an edge the scope can show, and a displayed rise time is always √(signal² + instrument²), so a slow scope makes every edge look slow. Probe ground-lead inductance resonates with the probe's tip capacitance and adds ringing that is not in your circuit. Learn the four numbers for the instruments on your bench and most instrumentation mistakes become impossible.
CORE CONCEPTThe multimeter: two resistors and a lot of arithmetic
As a voltmeter: a 10 MΩ resistor in parallel
A typical handheld digital multimeter presents about 10 MΩ in parallel with roughly a hundred picofarads on its DC voltage ranges. On almost every node you will ever probe this is invisible. On the nodes where it is not, Lesson 1.3's Thevenin arithmetic gives the answer immediately: the reading is Vth · Rin / (Rth + Rin).
Beacon's battery divider is 2 × 1 MΩ, so Rth = 500 kΩ. A 10 MΩ meter reads 10 / 10.5 = 95.2 % of the true midpoint — about 4.8 % low, or 88 mV on a 1.85 V node. That is large enough to matter and small enough to be mistaken for a calibration error. The meter is not wrong; it is reporting the voltage of the circuit that exists while the meter is attached. Where this really bites is when you use the meter to "check" a node and then compare against a firmware reading: two different loads, two different circuits, two different correct answers.
Resolution is not accuracy
A meter's digit count is resolution: a 4½-digit meter has 19,999 counts and will happily display 1.8503 V. Its accuracy is a separate specification, written in the form ±(0.5 % of reading + 3 counts). On a 1.8503 V reading with a 2 V range, that is 0.5 % × 1.8503 = 9.3 mV plus 3 counts × 0.1 mV = 0.3 mV, so about ±9.6 mV — the last two displayed digits are decoration. The source has a floor of its own, too: thermal noise is √(4kTRB), so Beacon's 500 kΩ battery divider generates about 91 µV RMS over a 1 MHz bandwidth. Averaging beats that down as √N and the meter's own ±9.6 mV does not, which tells you immediately which of the two is worth attacking. Two consequences. First, never report more precision than the accuracy specification supports. Second, a cheap meter's accuracy specification usually assumes a recent calibration and a stated temperature range, and drifts outside them.
As an ammeter: a resistor in series, and the number nobody reads
To measure current, a meter puts a shunt resistor in series with your circuit and measures the voltage across it. That voltage is called the burden voltage, and it is subtracted from your circuit's supply whether you like it or not. The shunt value differs enormously by range: on the amps range it is milliohms, on the milliamps range typically single-digit ohms, and on the microamps range typically hundreds of ohms to a few kilohms — because to resolve microamps you need a big enough voltage to measure. Good meters specify burden voltage explicitly, usually in millivolts per milliamp or as a full-scale figure; cheap meters do not, and you must measure it or infer it.
Put a meter on its µA range, with a shunt on the order of a kilohm, in series with Beacon's supply and try to capture a BLE burst. The burst wants 40 mA. Ohm's law says the burden voltage would be 40 V. What actually happens is that the rail collapses, the brown-out detector fires, the board resets, and you conclude the firmware has a bug in its radio code. Conversely, put the meter on the mA range with a 1 Ω shunt and try to read 8 µA of sleep current: the burden voltage is 8 µV, far below the meter's resolution, and you read zero or noise. One shunt cannot span this range, and autoranging does not save you — many meters break the circuit briefly while changing range, which resets the board at exactly the interesting moment.
Two more habits worth having
Four-wire measurement. When you measure a resistance below a few ohms, the leads and their contact resistance are a large fraction of the reading. A four-wire (Kelvin) measurement forces current through one pair and senses voltage with another, so the lead drop is excluded. This is how you measure a current-sense shunt, a trace, or a solder joint. Two-wire readings below about 1 Ω are approximately fiction.
True RMS, and what it is for. A cheap meter measures the average of a rectified AC waveform and scales it assuming a sine. Feed it a PWM signal, a switching regulator's ripple, or a chopped waveform and it reports a number with no physical meaning. A true-RMS meter integrates the square. For embedded work this matters most when you measure the current drawn by a duty-cycled load — and even a true-RMS meter's bandwidth on AC ranges is often only a few kilohertz, so it will not capture a 45 ms burst correctly either. The right tool for a duty-cycled current is not a meter at all.
EXAMPLEMeasuring 8 µA and 40 mA in one breath
Beacon's canonical numbers are a deep-sleep current of about 8 µA and an active window of about 45 ms at 40 mA average. The ratio between them is 5,000 : 1, and both must be measured on the same board, in the same minute, without disturbing it. This is the hardest routine measurement in low-power embedded work and it is worth walking through properly, because the arithmetic explains every tool people buy for it.
Why one shunt cannot do it
Suppose you insert a sense resistor in the battery lead and watch it with an oscilloscope. Pick the shunt for the sleep current first: to get a signal a scope can see above its own noise, you want tens of millivolts. At 8 µA, 10 kΩ gives 80 mV — comfortable. Now the radio fires: 40 mA through 10 kΩ would be 400 V, so in reality the rail collapses long before that and the board dies. Pick the shunt for the burst instead: 1 Ω gives 40 mV during the burst, fine — and 8 µV during sleep, which is below the noise floor of any general-purpose scope by a wide margin. There is no single value that works. That is not a limitation of your equipment; it is a property of the measurement.
The four things people actually do
- Measure the two regimes separately. Hold the board in deep sleep (a firmware test mode, or a jumper) and measure with a high shunt or a good meter on its µA range; then hold it transmitting continuously and measure with a low shunt. Combine arithmetically using the known duty cycle. Cheap, correct, and it misses transient behaviour entirely — the inrush at wake, the peak during a packet, the tail while the radio powers down.
- Auto-ranging current adapter. A small circuit that switches shunts automatically and outputs a scaled voltage into a scope or meter. Dave Jones's µCurrent is the well-known open design of this type. It solves the range problem and keeps the burden voltage low, at the cost of a range-switching artefact at the moment of transition.
- A dedicated power analyser. Instruments like the Nordic Power Profiler Kit II, Otii Arc, or Joulescope are built for exactly this: they range automatically in hardware, sample fast enough to resolve a burst, and integrate charge over a whole cycle so you get an average current rather than a picture you have to integrate by eye. If you are building battery products, one of these pays for itself in the first week.
- Scope on a sense resistor with a differential probe. Best transient detail, most setup, and the one that most often introduces its own errors — the shunt changes the supply impedance, the probe's ground reference matters, and you are integrating a waveform by hand.
For a duty-cycled product, "average current" is a derived figure: total charge per cycle divided by the cycle time. A scope trace of a burst gives you charge if you integrate it, which is what the maths function on any modern scope will do for you. The reason this framing matters is that it makes the comparison honest — a shorter burst at higher current can carry the same charge as a longer one at lower current, and only the charge shows up in the battery. Beacon's canonical "45 ms at 40 mA" is exactly this: not a constant 40 mA for 45 ms, but a composite window whose integral equals 1.8 mA·s of charge, or about 0.5 µAh per report. Multiply by 1,440 reports a day and you get 0.72 mAh a day, or 263 mAh a year — which against a 1000 mAh cell is the whole story, computed in one line.
HOW IT WORKSBandwidth, sample rate, memory, and the probe
Bandwidth is a rise-time filter
An oscilloscope's bandwidth is the frequency at which its response has fallen by 3 dB. For a scope with the classical single-pole-ish response, that corresponds to a rise time of:
tr(scope) ≈ 0.35 / BW
A 100 MHz scope has a rise time of about 3.5 ns; a 20 MHz one about 17.5 ns. Higher-end scopes with flatter, more brick-wall responses use a constant nearer 0.4 or 0.45, so check your vendor's figure if it matters. What you see on screen is the root-sum-square of the signal and the instrument:
tr(displayed) = √(tr(signal)² + tr(scope)² + tr(probe)²)
Put an ESP32 GPIO edge with a real rise time of about 3 ns onto a 100 MHz scope: √(3² + 3.5²) = 4.6 ns displayed. You are measuring the scope as much as the signal. Onto a 20 MHz scope: √(3² + 17.5²) = 17.8 ns — the reading is 99 % instrument. The standard rule follows: pick a bandwidth at least five times the highest frequency of interest, or equivalently at least 1/(3 · tr) of the edge you want to measure honestly.
Sample rate and memory depth are separate constraints
Bandwidth is analog; sample rate is what the digitiser does afterwards. Nyquist says you must sample above twice the highest frequency present, but a scope needs more than the theoretical minimum to render a waveform you can trust — vendors typically build sample rates four to five times their analog bandwidth, and with sine-interpolated reconstruction you want at least four or five samples across a rise time. With linear interpolation, more.
The constraint that surprises people is memory depth. The sample rate you actually get is memory depth divided by the capture window: a scope with 1 Mpoint of memory capturing 100 ms is running at 10 MSa/s no matter what its front panel claims. This is exactly the situation when you capture Beacon's 45 ms active window — you want the whole burst on screen, and the scope quietly drops its sample rate to fit. If you then zoom in on an edge, you are zooming into data that was never captured.
The probe is a capacitor on a stalk
A 1× probe presents roughly 100 pF at the tip (cable capacitance dominates) and is typically bandwidth-limited to something like 6 MHz. A 10× probe divides the signal by ten, which lets it present roughly 10 pF and reach hundreds of megahertz, at the cost of a ten-times-smaller signal into the scope's own noise. Low-capacitance probe tips reach a few picofarads. That capacitance is not a specification of the instrument — it is a component you have attached to the node, and it acts exactly like the strays of Lesson 1.4.
Worked example, and it is the one that catches people. Probe an I2C line pulled up by 4.7 kΩ. The bus's own rise time is set by the pull-up and the bus capacitance. Add a 1× probe's 100 pF and you have added a time constant of 4.7 kΩ × 100 pF = 470 ns, so a 10–90 % contribution of about 1 µs. At 400 kHz the whole bit period is 2.5 µs. You have not measured a marginal I2C bus; you have created one. Swap to a 10× probe and the added time constant drops to 47 ns and the bus behaves again. Also remember to compensate the probe against the scope's calibration square wave: an uncompensated 10× probe over- or under-shoots on low-frequency square waves and will make you doubt a perfectly good signal.
The ground lead is an inductor, and it rings
The probe's tip capacitance and the loop formed by the ground lead make a series LC circuit. A rule of thumb puts wire loop inductance near 1 nH per millimetre, so a 15 cm ground clip lead contributes on the order of 150–200 nH. Resonance:
fring = 1 / (2π √(Lgnd · Cprobe))
- 10× probe (10 pF) with a 15 cm ground clip (200 nH): f ≈ 113 MHz. On a 100 MHz scope this is right at the edge of visibility and appears as a soft overshoot.
- 1× probe (100 pF) with the same lead: f ≈ 36 MHz — squarely visible on any scope, and this is the classic "tens of megahertz ringing" that engineers chase for a morning.
- 10× probe with a 10 mm spring ground (about 12 nH): f ≈ 460 MHz, above most scopes' bandwidth, so it simply is not displayed.
The essential point: shortening the ground lead does not fix ringing, it stops you from adding ringing. Any oscillation whose frequency moves when you change the ground lead is yours. Any oscillation that stays put when you change probes, ground leads, and probe positions is the circuit's, and then it is worth designing a snubber for.
A mains-powered bench oscilloscope's probe ground is connected to earth through the mains cable. Clip it to a node that is not at earth potential and you have shorted that node to earth through the probe's ground lead — which can destroy the board, the probe, or the scope, and on a mains-referenced circuit can be lethal. Two safe approaches: use a differential probe, or float the circuit (battery power or an isolated supply) rather than the scope. Never lift the scope's earth pin. On Beacon this matters when you put a scope across a high-side shunt in the battery lead: neither end of that shunt is at the board's ground, so it needs two channels subtracted, or a differential probe, or the shunt moved to the low side where one end is ground.
SEE ITScope setup trainer — which artefacts are the signal and which are you
Pick a signal from Beacon's bench, then set up the scope to capture it. The trainer computes what the probe does to the node, what the bandwidth does to the edge, whether the ground lead's resonance is excited and visible, and whether the sample rate is enough to render what remains. The chart shows the true node waveform against what the screen would show, sampled and held. Every number is deterministic; the probe and ground-lead values are typical of the described hardware.
Things to try. Select the GPIO edge with the default 10× probe, 100 MHz, 1 GSa/s and a 20 cm ground clip: the displayed rise time is already about a third instrument, and the ground clip adds a visible overshoot. Switch to the spring ground and the overshoot vanishes without the signal changing at all — that is the morning of snubber design saved. Now select the I2C edge and switch to the 1× probe: the true rise time at the node changes, because 100 pF against a 4.7 kΩ pull-up is a real time constant you just added, and the bus that looked healthy now looks marginal. That one is not a display artefact — you genuinely broke the bus while measuring it. Next, select the buck switch node and drop the bandwidth to 20 MHz: the 1.5 MHz square wave still looks like a square wave, and every piece of information you were trying to see about the switching edges is gone. Finally take the BLE burst envelope, keep DC coupling and drop the sample rate to 1 MSa/s: perfectly adequate, because the envelope is slow — then switch to AC coupling and watch the measurement you needed, an absolute current level, disappear entirely.
IN PRACTICELogic analyzers, and the questions only a scope can answer
A logic analyzer samples many channels, but only as ones and zeros against a threshold. That trade buys channel count, capture depth, and — the real prize — protocol decoding: a good analyzer will show you I2C as addresses, ACKs, and bytes rather than as edges. For firmware bring-up this is transformative. When Beacon's SHT40 driver returns garbage, the analyzer tells you in seconds whether the address was wrong, whether the sensor NACKed, whether the read happened before the conversion finished, or whether the bus never moved at all.
| Question | Right tool | Why |
|---|---|---|
| Did the MCU send the right bytes to the right address? | Logic analyzer | Protocol decode, many channels, long captures |
| Did the sensor ACK, and when did it stretch the clock? | Logic analyzer | Decoders mark ACK/NACK and clock stretching explicitly |
| Are the bus edges reaching valid logic levels in time? | Oscilloscope | An analyzer sees only the threshold crossing; the shape is invisible to it |
| Why does the bus fail only when the radio transmits? | Oscilloscope | Analog coupling, rail droop, and glitch amplitude are analog phenomena |
| How much charge does the burst take? | Scope on a shunt, or a power analyser | It is an integral of an analog quantity |
| How long between the motion interrupt and the first advertisement? | Either — analyzer is easier | Long capture with a GPIO marker toggled by firmware |
Sample-rate advice for logic analyzers differs from scopes because you are only recovering ones and zeros: aim for at least four times the bus clock, and more if you care about locating an edge precisely. Beacon's I2C at 400 kHz is comfortable at 4 MSa/s and luxurious at 24 MSa/s, which is what an inexpensive USB analyzer offers. The channel-count trap is that many cheap analyzers reduce their sample rate as you enable more channels, so check what you are actually getting.
An I2C bus with pull-ups that are too weak for its capacitance has rounded rising edges. The analyzer's threshold is crossed eventually, so the analyzer decodes the transaction perfectly and reports that everything is fine. The receiver's input threshold is not the analyzer's threshold, and its setup time relative to the clock edge is not met, so the sensor occasionally samples the wrong bit — and only when the board is warm, or when the radio adds noise, or on the units with slightly higher bus capacitance. The analyzer says the bus is healthy on every capture you take. The scope shows a rise time that eats a third of the bit period, and the fix is a smaller pull-up resistor. The rule: an analyzer tells you what was transmitted, a scope tells you whether it should have been received. When a bus works "usually", reach for the scope.
One more habit that costs nothing and saves days: instrument your firmware with a GPIO. Toggle a spare pin high at the start of an operation and low at the end, and every instrument on the bench — analyzer, scope, power profiler — gains a trigger and a time reference. Beacon's active window measurement uses exactly this: a marker pin high for the whole wake, so the power profiler can integrate charge over precisely the interval the firmware defines rather than over an interval you eyeballed.
BEACONHow the canon's numbers were obtained, and what would have made them wrong
The canon states a deep-sleep current of about 8 µA and an active window of about 45 ms averaging 40 mA. Those are measurements, not datasheet sums (Lesson 1.6's table marks them so). Here is the method, and at each step the mistake that would have produced a plausible wrong answer.
Measure the board, not the development kit
Stages 1 to 3 prototype on an ESP32-C3-DevKitM-1, which carries the same ESP32-C3-MINI-1 module Beacon uses — that is why it is the honest prototyping platform. But the kit also carries its own LDO, a USB-serial interface, an LED, and pull resistors, and those dominate any sleep-current measurement by orders of magnitude. To measure the module you must break out its supply and measure there, or measure the finished Beacon board. The mistake: measuring the dev kit's USB current and reporting it as the product's sleep current, which produces a figure in the milliamps and an immediate conclusion that deep sleep does not work.
Disconnect the programmer before measuring sleep
A connected USB or debug interface holds pins at defined levels, powers internal domains, and on some parts prevents the deepest sleep states from being entered at all. Beacon's sleep measurement is taken with the board on battery only, no USB, no probe on any GPIO — because even a 10 MΩ scope probe on a high-impedance node can matter at the 8 µA scale, and a probe on a strapping pin (GPIO2, GPIO8, or GPIO9 on this part) can change boot behaviour outright. The mistake: a sleep current that is stable, repeatable, and three times too high because a probe was left clipped to a test point.
Range for the regime you are measuring, and mark the window in firmware
Sleep is measured with a high-value shunt or an instrument that ranges automatically; the active window is measured with a low-value shunt and a fast capture. The two are stitched together using a firmware marker GPIO that is high for exactly the wake period, so the active window's duration is defined by the firmware rather than estimated from a screen. The 40 mA figure is then the integral of current over that marked window divided by its length — an average, not a peak. The mistake: reading the peak off the screen and calling it the average, which for a bursty radio load overstates the charge by a factor of two or more and makes the battery projection pessimistic enough to trigger a redesign nobody needed.
Check the burden voltage against the brown-out threshold
Whatever shunt is in the battery lead drops I × R during the burst. At 40 mA a 1 Ω shunt costs 40 mV, which is harmless. A 10 Ω shunt costs 400 mV, which on a cell already down at 3.2 V pushes the regulator toward dropout and can trip the ESP32-C3's brown-out detector mid-transmission. The symptom is a board that resets only while being measured. The mistake: concluding the radio firmware is unstable when the instrument is the instability — and this is the specific reason the canon's numbers were taken with a low-burden instrument rather than a meter in series.
Repeat at temperature, and on more than one unit
Sleep current is leakage-dominated and roughly doubles every 8 to 10 °C, so a bench measurement at 22 °C says little about the 45 °C top of R9's environment. The pilot's 25 units (Stage 6) give the distribution that a single bench board cannot. The mistake: a single measurement on a single unit at room temperature, promoted to a specification — which is precisely the failure Lesson 1.6 opened with, arrived at from the measurement side rather than the datasheet side.
For every measured number in Beacon's requirements table, the record carries: the instrument and its range, the shunt value and resulting burden voltage, what was connected to the board, the firmware build and its sleep configuration, the ambient temperature, and how many units. That is roughly six lines per number. It sounds bureaucratic until the day a Rev B board reads differently and you need to know whether the board changed or the method did. Every experienced hardware engineer has lost a week to a measurement they could not reproduce, and every one of them writes this down afterwards.
WEAKNESSWhere this bites
The intuition that "noise is random, so a repeatable signal is real" is exactly backwards for instrumentation artefacts. Probe ringing is perfectly repeatable — it is a resonance, not noise. So is the rise time added by a probe's capacitance, the sag caused by a burden voltage, and the aliased frequency produced by an inadequate sample rate. All four look like solid, believable, reproducible measurements. The only reliable test is to change the measurement and see whether the result changes: swap probes, shorten the ground, change the sample rate, move the probe to a different point on the same net. A real signal is indifferent to all of these. This one habit separates people who debug hardware quickly from people who do not.
Scopes and meters both offer averaging, and it is genuinely useful for pulling a repetitive signal out of noise. It is also the fastest way to destroy exactly the evidence you need. A glitch that occurs on one cycle in a thousand averages to nothing. A brown-out that lasts 50 µs disappears into a meter's 200 ms integration window. Beacon's most likely field bug — a reset during the radio burst under a specific battery voltage — is invisible to every averaged measurement and obvious on a single-shot capture with a well-chosen trigger. Reach for averaging only when you already know the signal is repetitive, and reach for single-shot, peak-detect, or a long-memory capture when you are hunting something intermittent.
Two versions. First, ground: Lesson 1.3 showed that a shared return conductor carrying another circuit's current makes "ground" a different voltage in different places. Probing a signal with the ground clip attached at a distant test point measures the signal plus whatever KVL says is across the copper in between, and during a radio burst that can be millivolts to tens of millivolts. Ground the probe to the return of the circuit you are measuring, as close as the layout allows. Second, supply: measuring a rail at the regulator output tells you the regulator is fine; the part that browned out is at the other end of a trace and a via. Measure at the load's own pins. Beacon's 4-pogo test-jig pads (canon, 6.3) exist so that production test measures at defined points rather than wherever a probe fits.
KEY INSIGHTHow this connects
- Voltage, Current, Resistance, and Power (1.2) — burden voltage is Ohm's law applied to a resistor you did not intend to put in the circuit, and the arithmetic is identical.
- Kirchhoff's Laws and Circuit Analysis by Hand (1.3) — the meter's 10 MΩ against a 500 kΩ Thevenin source is the loading problem again, and the shared-ground error is KVL in the return path.
- Passive Components for Real (1.4) — a probe's tip capacitance behaves exactly like the strays that lesson catalogued, and the ground lead's inductance is the ESL story with a different name.
- Reading a Datasheet Without Getting Fooled (1.6) — this lesson supplies the numbers that no datasheet can give you, and the previous one tells you which those are.
- Breadboards, Prototyping, and the Bench (1.8) — the next lesson adds the breadboard's own parasitics to everything measured here, which is why some measurements should not be attempted on one at all.
- Debugging Firmware (3.9) — where the instruments here are combined with logs and JTAG into a method, rather than used one at a time on a hunch.
- Sleep Modes and the Power Budget (4.3) — where the 8 µA and 45 ms measurements are made for real and defended for 72 hours on the bench.
Everything in this course that claims a number eventually rests on somebody having measured it correctly. The discipline is small and it never changes: know what your instrument adds to the circuit, compute whether it matters before you connect it, and when a result surprises you, change the measurement before you change the design.
GLOSSARYKey terms introduced
- Burden voltage
- The voltage dropped across the shunt inside an ammeter, subtracted from the circuit under test. Ranges from microvolts on amp ranges to volts on microamp ranges, and is the reason a meter in series can stop a board from running.
- Input impedance
- The load a voltmeter or probe places on the node, typically 10 MΩ with about 100 pF for a handheld meter. Combines with the source's Thevenin resistance to set the reading error.
- Accuracy specification
- Usually of the form ±(x % of reading + n counts). Distinct from resolution, which is only how many digits are displayed.
- Bandwidth (−3 dB)
- The frequency at which a scope or probe's response has fallen by 3 dB. Related to rise time by tr ≈ 0.35/BW for a classical response, and combined with the signal as a root sum of squares.
- Sample rate and memory depth
- Two separate limits. The achieved sample rate is memory depth divided by capture time, so a long timebase silently reduces it regardless of the scope's headline figure.
- Aliasing
- Frequency content above half the sample rate folded down to a lower, entirely fictitious frequency. Looks like a real, stable signal, which is what makes it dangerous.
- Probe ground-lead inductance
- Roughly 1 nH per millimetre of lead, resonating with the probe's tip capacitance to add ringing at 1/(2π√(LC)). Not present in the circuit; present only while you measure it.
- Probe compensation
- Trimming a 10× probe's capacitance to match the scope input so that low-frequency square waves are reproduced flat. An uncompensated probe over- or undershoots on every square wave you look at.
ACTIVE RECALLCheck your understanding
Answer in your own words first, then reveal.
1You put a handheld meter in series with Beacon's battery to measure sleep current. It reads 7.9 µA, which matches the canon. Then the board wakes to transmit and the meter reads nonsense and the board resets. Explain both observations and give two fixes.›
Both come from the same shunt. On the µA range the meter's shunt is large — typically hundreds of ohms to a few kilohms — which at 8 µA produces a burden voltage of a few millivolts, small enough that the sleep reading is honest. When the radio fires and the board wants 40 mA, that same shunt would need to drop volts. What actually happens is that the supply collapses, the brown-out detector trips, and the board resets, which also means the meter never sees a steady 40 mA to display. If the meter is autoranging it will try to switch ranges, and many meters open the circuit briefly while doing so, guaranteeing the reset. Fix 1: measure the two regimes separately, with the board held in each state by a firmware test mode, and combine them arithmetically using the known duty cycle. Fix 2: use an instrument built for the job — an auto-ranging current adapter or a power profiler — which switches shunts in hardware fast enough to keep the burden voltage low in both regimes and integrates charge over the whole cycle. Both are better than any single resistor, because the problem is 5,000 : 1 of dynamic range, not a shortage of resolution.
2You measure an ESP32-C3 GPIO edge as 18 ns on a 20 MHz scope and 4.6 ns on a 100 MHz scope. Which is right, and what is the actual rise time?›
Neither is the signal; both are correct readings of signal-plus-instrument. Using tr(scope) = 0.35/BW: the 20 MHz scope contributes 17.5 ns and the 100 MHz scope 3.5 ns. Back out the signal from each with tr(sig) = √(tdisplayed² − tscope²). From the 20 MHz reading: √(18² − 17.5²) = √(324 − 306) = 4.2 ns, and notice how badly conditioned that subtraction is — a one-nanosecond error in the reading moves the answer by several nanoseconds. From the 100 MHz reading: √(4.6² − 3.5²) = √(21.2 − 12.3) = 3.0 ns, which is well conditioned because the instrument is the smaller term. The actual rise time is about 3 ns, and the 100 MHz measurement is the one to trust. The general lesson is not "use a faster scope" but an instrument contribution comparable to the signal makes the subtraction useless. If you need to measure a 3 ns edge properly, use a scope where 3 ns is a large multiple of the instrument's own rise time — which is the origin of the five-times-bandwidth rule.
3A colleague shows you 30 MHz ringing on Beacon's 3.3 V rail during the BLE burst and proposes adding a series ferrite bead and a snubber. What do you check first, and what would convince you the ringing is real?›
Check the probe before the board. Thirty megahertz is squarely in the range produced by a 1× probe's roughly 100 pF resonating with a 15–20 cm ground clip's 200–300 nH: 1/(2π√(250 nH × 100 pF)) ≈ 32 MHz. So the first three actions are free and take a minute. (a) Change the ground lead. Swap the clip for a spring ground. If the ringing frequency moves, it is the probe. Real circuit resonances do not care what you ground with. (b) Change the probe. Move from 1× to 10×; the frequency should roughly triple if it is the probe's LC. (c) Move the probe along the same net. A real rail resonance appears everywhere on the net with amplitudes that vary sensibly; a probe artefact follows the probe. What would convince me it is real: the frequency stays put across two probes, two ground lengths, and two probe points; it correlates with a plausible physical mechanism (an inductance and a capacitance you can identify on the schematic, such as the bulk capacitor's ESL with the decoupling); and it appears in a second, independent measurement — for example, as a disturbance on a nearby signal, or on the current waveform through the shunt. Only then is a snubber worth designing, and by then you also know which L and which C to damp.
4Beacon's SHT40 driver works on the bench and fails on about one board in five, always reporting a NACK on the first read after a cold boot. A logic analyzer says every transaction is well-formed. What is your plan?›
The analyzer's verdict is the clue, not a dead end: it means the bits are correct at the analyzer's threshold, so the fault is either analog or temporal. Three hypotheses and a measurement for each. (a) Rise time. Put a scope on SCL and SDA and measure the rise time against the bit period. If the pull-ups are marginal for the bus capacitance, the edge crosses the sensor's own input threshold late; the analyzer, whose threshold and timing are different, sees a valid bit. Cheap boards vary in bus capacitance, which explains why some units fail. Fix: smaller pull-ups, computed against the I2C specification's rise-time limit for the bus speed. Remember also that your probe adds capacitance — measure with a 10× probe, not a 1×, or you will manufacture the symptom. (b) Power-on timing. The SHT40 has a soft-reset and start-up time after its supply rises; a read issued before the sensor is ready NACKs. Measure with a marker GPIO set at the first line of firmware and compare to the rail's rise on the scope. Fix: honour the datasheet's power-up time, using the maximum rather than the typical. (c) Rail integrity at cold boot. If the regulator's start-up is slow or the rail dips as the radio initialises, the sensor may brown out and reset. Single-shot capture on the rail, triggered by the marker pin. Note that all three are analog or timing questions, and the analyzer is blind to all three — which is exactly the boundary between the two instruments.