Regulators: LDO vs Buck vs Boost
An LDO turns excess voltage into heat. A buck turns it into current. On a battery, that is the whole argument.
🎯 By the end you'll understand
What a linear regulator physically does with the voltage it throws away, why its efficiency is simply Vout/Vin, and what dropout and quiescent current cost you. How a buck converter moves energy through an inductor instead of burning it, why its efficiency curve is high in the middle and collapses at light load, and why a modern part like the TPS62840 (60 nA quiescent) changes the answer for sleepy devices. When a boost converter belongs in the design. And the discipline that matters more than any of the parts: choose a regulator for where the device spends its time, not for its peak — worked through Beacon's move from the XC6220 LDO on the breadboard to the TPS62840 on the PCB, with numbers.
KEY INSIGHTEvery regulator is a deal about where the surplus goes
The forum said "use a buck, it's 95 % efficient." The board came back drawing 35 µA asleep instead of 8. The buck was 95 % efficient — at 100 mA. At 8 µA it was 20 % efficient, and worse than the LDO it replaced.
A battery gives you one voltage that drifts from 4.2 V to 3.0 V. Your chips want another — 3.3 V, steadily. A regulator closes that gap, and there are only two honest ways to do it. A linear regulator (an LDO is the low-dropout kind) puts a variable resistor in series and drops the surplus across it as heat. A switching regulator chops the input on and off and uses an inductor and capacitor to average the chops into a lower (buck) or higher (boost) voltage, converting surplus voltage into extra current instead of heat. The linear part is simple, quiet, and wastes exactly Vin − Vout per amp. The switcher is efficient and noisy, and it has a fixed cost of running — gate charge, control logic, timing — that does not shrink when the load does.
Efficiency is not a number a regulator has; it is a curve against load current. Read that curve at the two points where your device lives — asleep (microamps, 99.9 % of the time) and awake (tens of milliamps, 0.1 % of the time) — and weight each by the fraction of the battery's charge that flows through it. For Beacon, most of the charge flows in the active bursts, but the sleep floor is where regulators differ by a factor of ten. The right regulator is the one that wins at the point that dominates the ledger, and you only know which point that is after Lesson 4.3's budget. This lesson gives you the curves.
HOW IT WORKSThe LDO: a transistor that dissipates the difference
Inside a low-dropout regulator there are three things: a voltage reference (typically a bandgap, about 1.2 V), an error amplifier, and a pass element — on modern parts a PMOS transistor between input and output. The error amplifier compares a divided-down copy of the output with the reference and drives the PMOS gate so that the output sits exactly where it should. The PMOS is operating as a controlled resistor: if the load draws more, the amplifier turns it on harder; if the input rises, it turns it off a little. All of the current that reaches the load passes through that resistor, and the voltage across it is Vin − Vout. The power it dissipates is therefore
Ploss = (Vin − Vout) × Iload + Vin × IQ
and the efficiency, ignoring the small quiescent term, is just η = Vout / Vin. That is the whole thermodynamics of a linear regulator: the current in equals the current out, so the ratio of powers is the ratio of voltages. From a fresh LiPo at 4.2 V into a 3.3 V rail you get 79 %; at the 3.7 V plateau, 89 %; at 3.5 V, 94 %. The LDO is most efficient when the battery is nearly empty, which is a strange and useful fact.
Three datasheet numbers describe an LDO, and each hides a trap.
Dropout voltage
The minimum Vin − Vout at which the part still regulates. It is the pass transistor's on-resistance times the load current, so it scales with load: a part whose headline dropout is a couple of hundred millivolts at its rated amp will drop only a few tens of millivolts at Beacon's 40 mA. Below dropout, the LDO does not "fail" — it just passes the input through, minus that small drop, and the output tracks the battery downward. On the discharge curve from Lesson 4.1, this is the line that strands the last few percent.
Quiescent current, IQ
What the regulator draws with no load: the reference, the error amplifier, the divider. It flows from the battery all the time. For a 3.3 V LDO in the XC6220's class it is in the single-digit microamps — invisible next to a 40 mA burst, but it is a full doubling of an 8 µA sleep floor. Some LDOs quote "ground current" that rises with load; read the curve, not the headline.
PSRR and noise
Power-supply rejection ratio says how much input ripple reaches the output, in dB, as a function of frequency. LDOs are excellent at low frequency (60–80 dB) and fall off above tens of kHz — exactly where a buck's switching ripple sits. That is why sensitive analog rails are sometimes fed by a buck followed by an LDO: the buck does the efficiency, the LDO does the cleaning, and you accept a few hundred millivolts of headroom as the price. Beacon's sensors are digital and tolerant; it does not need this.
HOW IT WORKSThe buck: an inductor that trades voltage for current
A buck converter is a switch, an inductor, a capacitor, and a controller. With the switch closed, the inductor sees Vin − Vout across it and its current ramps up, storing energy in its magnetic field while also feeding the load. With the switch open, the inductor's current cannot stop instantly (that is what inductance means, from Lesson 1.4), so it keeps flowing through a second path — a diode, or on any modern part a second synchronous MOSFET — and the inductor now sees −Vout, ramping its current down. Average the switch node over a cycle: Vout = D × Vin, where D is the fraction of the time the switch is on. Nothing in that loop dissipates by design; the only losses are the switches' on-resistance, the inductor's winding resistance and core, and the energy it takes to charge and discharge the MOSFET gates every cycle. So a buck delivering 40 mA at 3.3 V from a 3.8 V battery draws about 3.3 × 0.040 / (3.8 × 0.93) ≈ 37 mA from the battery. An LDO would draw 40 mA. The difference at this operating point is small — smaller than the forum argument suggests.
Because the inductor current ramps up and down, the output capacitor sees a triangular current and the output has a small triangular ripple: ΔV ≈ ΔIL / (8 × fsw × C) in continuous conduction. At a 1.8 MHz-class switching frequency with a 10 µF ceramic, ripple at Beacon's currents is a few millivolts — until DC-bias derating turns your "10 µF" 0603 into 4 µF at 3.3 V, as Lesson 1.4 warned. At light load the converter changes behavior entirely, which is the next section.
The switching also produces conducted and radiated noise at fsw and its harmonics. The high-frequency current loop — input capacitor, high-side switch, low-side switch, back to the capacitor — must be physically tiny on the PCB, or it becomes an antenna. That is a Lesson 5.6 problem; note here only that a buck is not a drop-in for an LDO on a breadboard, which is exactly why Beacon prototyped with the XC6220 and moves to the TPS62840 only on the PCB.
CORE CONCEPTThe light-load collapse, and the parts that fixed it
Here is the failure mode from the opening. A buck has a fixed cost per switching cycle: charging the MOSFET gates, running the oscillator and comparators, biasing the reference. At 100 mA that overhead is a rounding error against the power delivered. At 10 µA, delivered power is 33 µW, and a controller that burns 30 µA of quiescent current is spending 110 µW to deliver 33 — an efficiency of about 20 %. Two things save modern parts.
The first is PFM, pulse-frequency modulation. Instead of switching at a fixed 1.8 MHz regardless of load, the controller switches only when the output droops below a threshold, delivering one burst of energy and then going quiet. At a 10 µA load the pulses become sparse — kilohertz or slower — and the switching overhead scales down with the load. The cost is larger output ripple (tens of millivolts, at a frequency that wanders with load) and, occasionally, audible noise from ceramics if the pulse rate lands in the audio band.
The second is quiescent current engineered down to nanoamps. The TPS62840 is the reference example: its datasheet quiescent current is 60 nA, and it holds useful efficiency down into single-microamp loads, where an older buck is in single-digit percent. This is not a minor spec improvement; it is what makes a buck viable at all for a device that sleeps at 8 µA. The part also accepts 1.8 V to 6.5 V at its input, which is the "reach" argument from Lesson 4.1 — it keeps regulating well below where a 3.3 V LDO has dropped out.
Skip the headline efficiency. Find the efficiency-vs-load graph, check that its x-axis goes down to at least 10 µA (if it stops at 1 mA, the vendor is hiding something), read the value at your sleep current, and read the quiescent-current line in the electrical characteristics table — noting whether it is specified with the output in regulation ("IQ, non-switching" vs "operating"). Then find the minimum input voltage. Those four numbers decide the part for a sleepy device; everything else is layout.
SEE ITRegulator picker: efficiency at the points that matter
Pick a regulator and set Beacon's profile: the load on the 3.3 V rail while asleep and while awake, how long the active window lasts, and how often it happens. The widget reads each regulator's curve at both operating points, converts rail current into battery current, and projects battery life — on paper, and with the canon's 6× real-world derating against the six-month requirement. Start by choosing the "generic buck", because that is what a search for "3.3 V buck module" returns.
Three experiments worth running. Compare the LDO and the TPS62840 at the default profile: the active-point efficiencies are a few percent apart, and the sleep-point ones are not — that gap is most of the difference in derated life. Then take the report period out to 10 minutes and watch the sleep point take over the ledger completely; now the generic buck is a catastrophe while the LDO, with nothing to power between reports, still clears R1. Finally, push the active load to 150 mA at a 45 ms window and see the LDO's Vout/Vin penalty finally matter — which is the profile of a device that is not Beacon.
COMPARISONBoost, buck-boost, and choosing by profile
A boost converter is the buck's mirror: the inductor charges from the input with the switch closed, then dumps its energy into the output through a diode or synchronous switch when it opens, so Vout ≥ Vin. You need one when the source is below the rail — a single 1.5 V alkaline cell, two AA cells at 2.0–3.2 V, a 3.2 V LiFePO4 feeding a 3.3 V rail near end of life, or a supercapacitor. Modern ultra-low-IQ boosts (TI's TPS61099 family is the usual example) reach sub-microamp quiescent current and efficiencies in the high 80s at milliamp loads, so the boost itself is not the problem. The problems are structural: a boost's input current is higher than its output current (Iin ≈ Iout × Vout / (Vin × η)), and a plain boost cannot disconnect its output — the inductor and diode form a DC path from input to output even when the converter is off, unless the part has "true shutdown".
A buck-boost handles a source that straddles the rail — a LiPo feeding 3.3 V would need one only if you insisted on regulation all the way down to 2.5 V. Beacon does not: below ~3.0 V the protection IC is about to disconnect the cell anyway, and the TPS62840 in pass-through delivers Vin minus a small drop, which the ESP32-C3 tolerates down to its own 3.0 V minimum. A buck-boost would cost efficiency and money for a few percent of the curve the cell should not be discharged into.
| Topology | Efficiency at 40 mA (3.8 V → 3.3 V) | Efficiency at 8 µA | Quiescent | Noise | Choose when |
|---|---|---|---|---|---|
| LDO (XC6220 class) | ≈ Vout/Vin = 87 % | ≈ 40–50 % (IQ dominates) | single-digit µA | None (cleans ripple) | Prototypes, noisy-sensitive rails, small Vin − Vout, cost-critical boards that are rarely asleep |
| Buck, modern (TPS62840) | ≈ 93 % | ≈ 80–85 % | 60 nA | Switching ripple, PFM at light load | Battery devices that sleep; Vin well above Vout; you control the layout |
| Buck, generic (IQ 30 µA class) | ≈ 94 % | ≈ 20 % | tens of µA | Switching ripple | Always-on loads above a few mA; never for a sleepy device |
| Boost (TPS61099 class) | ≈ 88–90 % (from 2.8 V) | ≈ 75–80 % | sub-µA | Switching ripple; input current higher than output | Source below the rail (AA cells, supercaps, LFP near empty) |
The discipline underneath the table: write down the load profile first. For each state the device can be in, note the current and the fraction of time. Multiply, sum, and you have the average charge per hour in each state. Then, and only then, read each candidate's curve at those states. A device that runs a motor 30 % of the time wants a different regulator than one that sleeps 99.9 % of the time, even if their peak currents are identical. Engineers who choose by peak get the generic buck; engineers who choose by profile get the TPS62840 — or, for a device that is never asleep, cheerfully keep the LDO.
BEACONBeacon: from the XC6220 on the breadboard to the TPS62840 on the PCB
Beacon has been running on an XC6220 LDO since Stage 1. That was the right choice for a breadboard: a three-terminal part with two capacitors, no inductor, no layout-sensitive loop, and no switching noise to confuse the first current measurements. Now that the profile is known, the PCB gets the TPS62840. Here is the argument in the canon's numbers.
| Operating point | Rail load | XC6220-class LDO → battery | TPS62840 → battery |
|---|---|---|---|
| Deep sleep (99.93 % of the time) | ≈ 8 µA | 8 µA + IQ ≈ 16 µA | ≈ 8 µA total (the buck's 60 nA and its light-load loss are inside the canon's measured 8 µA) |
| Active window (45 ms per minute) | ≈ 40 mA | ≈ 40 mA (η ≈ 87 % at 3.8 V, as heat) | ≈ 37 mA (η ≈ 93 %) |
| Active contribution to average | 40 mA × 45 / 60 000 = 30 µA | 37 mA × 45 / 60 000 ≈ 28 µA | |
| Average from the battery | ≈ 46 µA | ≈ 36–38 µA (canon: 38 µA measured) | |
| Paper life on 1000 mAh | ≈ 21,700 h ≈ 2.5 years | ≈ 26,000 h ≈ 3 years | |
| Derated 6× (the canon's demonstration rule) | ≈ 5 months — misses R1 | ≈ 6 months — meets R1 | |
| Usable discharge window | 4.2 → ~3.4 V (~95 % of the cell) | 4.2 → 3.0 V (~99 %), regulates to 1.8 V in |
Notice where the win comes from. The active-point efficiency difference is worth about 2 µA of average. The sleep-point difference — an LDO's quiescent current versus 60 nA — is worth about 8 µA, four times more. And the discharge-window difference is worth another five or six weeks that never appear in the average-current arithmetic at all. The forum's "95 % efficient" argument is the least important of the three.
An input capacitor and an output capacitor of the values in its typical application (ceramic, sized after DC-bias derating), a small inductor with a saturation current comfortably above Beacon's peaks, a tight switching loop in the layout (5.6), and a decision about the light-load mode. It also offers a way to pause switching briefly while a noise-sensitive measurement is taken — worth checking in the datasheet if you ever hang a precision analog sensor on this rail. Beacon's SHT40 and LIS2DH12 are digital; the default configuration is fine.
One honest caveat about the canon's 8 µA sleep number: it is the total board, measured at the battery, with the buck in place. Of that, the module in deep sleep is about 5 µA per its datasheet, the LIS2DH12 in low-power 25 Hz mode is a few microamps, the SHT40 idles at a fraction of a microamp, and the buck's own quiescent draw plus its light-load conversion loss make up most of the remainder. Lesson 4.3 takes that ledger apart line by line, and shows what happens when a 1.8 µA leak is added to it.
WEAKNESSWhere this bites
The single most common power-system mistake on battery devices: the engineer sizes for the burst current (because that is what the bench supply's ammeter shows during development) and chooses a switcher whose curve is superb at 100 mA and unspecified below 1 mA. The device then sleeps at 30–40 µA instead of 8, and nobody notices until the field units die in a third of the projected time. If the datasheet's efficiency graph does not reach your sleep current, the part is not a candidate.
Below dropout an LDO passes the input straight through, minus a small drop, with no regulation at all. On a LiPo this is benign — the rail sags toward 3.0 V along with the cell, and the ESP32-C3 tolerates that. On a 5 V input feeding a 3.3 V rail it is irrelevant. But on a design where Vin can rise above the LDO's absolute maximum, or where a downstream part cannot tolerate the raw input, "it just passes through" is a fault, not a feature. Check what every rail sees when the regulator is out of regulation, in both directions.
A buck's high-frequency loop through jumper wires is a loop antenna with centimetres of inductance; the ripple, EMI, and sometimes the stability are nothing like the PCB's. Sleep-current measurements taken that way include the converter misbehaving. Beacon's sequence — LDO on the breadboard, buck on the PCB, re-measure — is not caution for its own sake; it is the only way to get a sleep number that means anything.
KEY INSIGHTHow this connects
- Batteries: Chemistry, Capacity, and the Lies of mAh (4.1) — the regulator's minimum input voltage is the cutoff line on the discharge curve; the buck's 1.8 V reach is what turns "95 % usable" into "99 %".
- Passive Components for Real: Resistors, Capacitors, Inductors (1.4) — the buck's inductor stores energy exactly as described there, and its "10 µF" ceramic output capacitor is subject to the DC-bias derating that lesson warned about.
- Sleep Modes and the Power Budget (4.3) — the sleep-point efficiency you read here becomes one line in the budget that decides Beacon's six months.
- Routing: Traces, Current, and Return Paths (5.6) — the buck's hot loop is the first thing you will place and route on the PCB, before anything else is allowed to move.
Regulator choice is where the battery lesson and the sleep lesson meet. You now have the curves; Lesson 4.3 gives you the time-weighting that turns them into a verdict.
GLOSSARYKey terms introduced
- LDO (low-dropout regulator)
- A linear regulator whose pass element (usually a PMOS) can regulate with only a small input-output difference. Efficiency ≈ Vout/Vin; the surplus becomes heat.
- Dropout voltage
- The minimum Vin − Vout at which an LDO still regulates; equals the pass element's on-resistance times the load current, so it shrinks at light load.
- Quiescent current (IQ)
- The current a regulator draws for its own operation with no load. Flows from the battery continuously; a direct addition to the sleep floor. 60 nA on the TPS62840; single-digit µA on a small LDO; tens of µA on older switchers.
- Buck converter
- A step-down switching regulator: a switch, an inductor, and a capacitor deliver Vout ≈ D × Vin, with input current ≈ D × output current. Losses are switching and conduction, not the voltage difference.
- PFM (pulse-frequency modulation)
- A light-load mode in which a switcher pulses only when the output droops, so switching losses scale down with load. Raises ripple; essential for microamp efficiency.
- Boost converter
- A step-up switching regulator, used when the source is below the rail. Input current exceeds output current; a plain boost has a DC path from input to output even when off unless it has true shutdown.
- PSRR
- Power-supply rejection ratio: how much input ripple an LDO removes, in dB, versus frequency. High at low frequency, falling above tens of kHz — where switcher ripple lives.
- Load profile
- The set of (current, fraction of time) pairs describing how a device spends its life. The only correct basis for choosing a regulator; peak current alone is not.
ACTIVE RECALLCheck your understanding
Answer in your own words first, then reveal.
1Derive the efficiency of an LDO from first principles, and explain why it is highest when the battery is nearly empty.›
All the load current passes through the pass transistor, so Iin = Iout (plus a small quiescent current). Power in is Vin × I, power out is Vout × I, so η = Vout/Vin. As a LiPo discharges from 4.2 V to 3.4 V, the ratio for a 3.3 V rail climbs from 79 % to 97 %, because there is less surplus voltage to burn. The LDO is least efficient on a full battery — which is also why its heat problem, if it has one, appears right after charging.
2A buck converter is "95 % efficient" on its datasheet's front page. Under what conditions could it still be a worse choice than an LDO for Beacon?›
When its quiescent and per-cycle switching overhead is large relative to the sleep load. At 8 µA, a converter with 30 µA of quiescent current delivers 33 µW while consuming over 100 µW — around 20 % efficient — and draws roughly 35 µA from the battery in sleep, versus about 16 µA for an LDO with single-digit-µA IQ. Since Beacon sleeps 99.9 % of the time, that sleep-point loss dominates the ledger and the "95 %" figure (read at 100 mA) is irrelevant. The fix is a part engineered for light load — the TPS62840's 60 nA IQ and PFM operation — not a different topology.
3List the three separate reasons the TPS62840 beats the XC6220 for Beacon, and rank them by how many microamps (or weeks) each is worth.›
(1) Sleep quiescent current: 60 nA versus single-digit microamps — worth about 8 µA of average current, the largest item. (2) Discharge-window reach: the buck regulates down to a 1.8 V input, recovering the ~4 % of the cell below the LDO's ~3.4 V cutoff — worth roughly 40 mAh, five or six weeks at 38 µA, and invisible in the average-current arithmetic. (3) Active-point efficiency: ~93 % versus ~87 % at 3.8 V — worth only about 2 µA of average, because the active window is 45 ms per minute. The popular argument is the smallest one.
4Why did Beacon prototype with an LDO on the breadboard rather than the buck it will ship with?›
Because a buck's high-frequency switching loop must be physically small to behave, and a breadboard's jumper wires turn it into a loop antenna with centimetres of stray inductance — producing ripple, EMI, and instability that have nothing to do with the final PCB. Sleep-current measurements taken that way would be measuring the converter's misbehavior rather than the design. An LDO has no such loop, needs only two capacitors, and adds a predictable single-digit-µA quiescent draw that can be subtracted. The buck goes on the PCB, where the layout can be controlled, and the sleep current is then re-measured.