Class-D Amplifier

Why a Class-D amplifier is not a louder Class-AB

Every linear audio amplifier — Class-A, Class-B, Class-AB — works by acting as a controllable resistor between the power supply and the loudspeaker. The output transistor is held part-way open, and the music is the moving voltage across the speaker terminals. The problem is what happens to the rest of the supply voltage: it is dropped across the transistor as heat. When a Class-AB amplifier delivers 10 V into an 8 Ω speaker from a 35 V rail, the transistor is dissipating 25 V × 1.25 A ≈ 31 W to deliver only 12.5 W of audio. The amplifier is more space heater than amplifier.

A Class-D amplifier refuses to play that game. Its output transistors are never held part-way open. They are switches: either fully closed (saturated, near-zero voltage across them, so V×I loss is tiny) or fully open (cut off, near-zero current, so V×I loss is tiny). A switch that is either fully on or fully off dissipates almost nothing. The trick is that you cannot represent an analog audio signal with a switch that only knows two states — so Class-D first converts the audio into the timing of those switch events using pulse-width modulation, and then reconstructs the analog signal afterward with an LC filter. The transistors only ever live in the two efficient states; the information lives in when they switch, not in how far they open.

That single move — encode in time instead of amplitude — is what takes a 200 W amplifier from dissipating 150 W of heat to dissipating 20 W. It is the reason a touring PA rack can deliver 20 kW from a 2U chassis, the reason a car subwoofer amplifier survives in a sealed trunk, and the reason your phone can be loud on a 3.7 V battery. The "D" does not stand for "digital" — it is just the next letter after the Class-A/B/C lineage — but the spirit is right: Class-D treats audio as something to be sampled in time rather than tracked in voltage.

Encoding audio as PWM — the comparator and the triangle

The heart of a Class-D amplifier is a single comparator and a single high-frequency reference waveform. The reference is usually a triangle wave (sometimes a sawtooth) running at a fixed carrier frequency f_sw, typically 250 kHz to 1 MHz — at least ten times the top of the audio band so the two never overlap. The comparator does exactly one thing, continuously:

if  V_audio(t) > V_triangle(t)   →   output HIGH
if  V_audio(t) < V_triangle(t)   →   output LOW

Because the triangle sweeps top-to-bottom many times per audio sample, the comparator flips high and low once per carrier cycle, and the fraction of each cycle spent high — the duty cycle D — tracks the instantaneous audio amplitude. This is natural-sampling PWM. With the audio at silence (mid-rail), the comparator spends half its time high: D = 50%. A loud positive peak pushes the audio above most of the triangle, so D climbs toward 100%. A loud negative peak drops D toward 0%. The relationship for a bridge output is linear:

D(t) = 0.5 · ( 1 + V_audio(t) / V_peak )

V_out(avg) = V_supply · ( 2·D − 1 )    (for a half-bridge, ±V_supply rails)

The audio is now entirely contained in the time-varying duty cycle of a two-level square wave. No amplitude information survives in the pulse height — that is fixed at the rail voltage. It all lives in pulse width. That is the central encoding step and the reason the architecture is sometimes called a switching amplifier or a PWM amplifier.

The switching output stage — half-bridge and full-bridge

The logic-level PWM cannot drive a speaker directly; it needs power. A gate driver buffers it up to the gate-charge currents (often several amps for tens of nanoseconds) needed to switch power MOSFETs quickly. The MOSFETs form a bridge:

• Half-bridge (single-ended). Two MOSFETs in series between the +rail and −rail, output taken from the midpoint. Needs a split supply (±V) and produces an output that swings ±V_supply. One LC filter per channel. Simplest, but the supply must source and sink, and DC offset must be managed.
• Full-bridge (BTL — bridge-tied load). Two half-bridges, the speaker connected between their two midpoints, driven in anti-phase. Runs from a single supply (0 to +V), doubles the voltage swing across the load (so 4× the power into the same impedance), and cancels even-order distortion. This is what almost every modern stereo and automotive Class-D chip uses.

The MOSFET choice is dominated by two figures of merit: on-resistance R_DS(on), which sets the conduction loss I²·R_DS(on) while the device carries speaker current, and gate charge Q_g, which sets the switching loss every time the device turns on or off. The two trade against each other — a bigger MOSFET has lower R_DS(on) but higher Q_g — so the optimum die size depends on the switching frequency. This is exactly why you cannot simply crank f_sw to 5 MHz for a "cleaner" PWM: switching loss scales with frequency, and at some point the switching loss you add exceeds the audio quality you gain.

Reconstructing the audio — the LC filter

The bridge output is a violent rail-to-rail square wave switching hundreds of thousands of times a second, with the music hidden in its duty cycle as the low-frequency average. To recover the audio you need to take that running average — and a second-order LC low-pass filter is the cheap, lossless way to do it. A series inductor and a shunt capacitor form a filter with corner frequency:

f_c = 1 / ( 2π · √(L·C) )

Example:  L = 15 µH,  C = 0.68 µF
f_c = 1 / (2π · √(15e-6 · 0.68e-6)) ≈ 49.8 kHz

That corner sits comfortably above the 20 kHz audio ceiling (so it passes all the music flat) and well below a 400 kHz carrier (so it kills the carrier by roughly 40 dB — a second-order filter rolls off at 12 dB/octave, and 400 kHz is ~3 octaves above 50 kHz, giving ~36 dB of attenuation plus margin). The filter is lossless in principle: the inductor and capacitor store and return energy rather than burning it, which is why the efficiency gain survives all the way to the speaker. The catch is that the filter's behavior depends on the load it is terminated into — the speaker. A speaker is not a flat 8 Ω resistor; its impedance rises at resonance and at high frequency. That interaction can cause a few decibels of frequency-response ripple, which is the single biggest reason high-fidelity Class-D designs close a feedback loop after the filter (post-filter feedback) to flatten the response no matter what the speaker does.

The efficiency budget — where the watts actually go

It is worth being concrete about why "90% efficient" is true. In a Class-D amplifier delivering output power P_out, the losses break into three buckets:

Loss mechanism	Formula	Scales with	Dominant when
Conduction loss	I_load² · R_DS(on)	Output current²	High output power
Switching loss	½ · V · I · (t_on+t_off) · f_sw	Carrier frequency	High f_sw, light load
Gate-drive loss	Q_g · V_gs · f_sw	Carrier frequency	Idle / quiet passages
Filter / parasitic	I² · (DCR + ESR)	Output current²	High output power

Consider a TI TPA3255 driving 150 W into 4 Ω. The speaker current is √(150/4) ≈ 6.1 A RMS, ~8.7 A peak. With R_DS(on) ≈ 80 mΩ per MOSFET and two in the conduction path, conduction loss is roughly I²·2·R_DS(on) ≈ 6.1² × 0.16 ≈ 6 W. Switching and gate-drive loss at 400 kHz add another ~8 W. Total loss ~15 W to deliver 150 W: efficiency ≈ 150 / 165 ≈ 91%. A Class-AB amplifier delivering the same 150 W into 4 Ω would dissipate on the order of 100–150 W of heat — needing a heatsink the size of the rest of the amplifier. That difference is the entire commercial case for Class-D.

Note the asymmetry: switching and gate-drive losses are nearly constant regardless of how loud you play, so at idle a Class-D amplifier still burns a few watts shuffling gate charge. This idle loss is why "efficiency" curves peak near full power and droop at low output, and why phone amplifiers go to elaborate lengths (clock-gating, lowering f_sw at low levels) to claw back milliwatts of standby battery life.

Modulation schemes — AD, BD, and filterless

• AD modulation (two-level). The two bridge outputs are exact complements; the differential output swings between +V and −V. Simple, but at idle the full carrier appears across the speaker and the filter inductors carry large ripple current, wasting power and radiating EMI.
• BD modulation (three-level). The two half-bridges are switched so the differential output takes three states: +V, 0, and −V. At idle both sides are in phase, so the differential voltage is zero and there is no ripple current — much lower idle loss and EMI. This is the default in modern parts, at the cost of more complex modulation logic.
• Filterless modulation. For tiny in-device speakers (phones, laptops) where the trace to the speaker is short and the speaker's own inductance does the averaging, the bulky LC filter is dropped entirely. The modulation is shaped so that the common-mode carrier cancels and the speaker sees only audio. Saves board area and cost, but only works for low power into a nearby, inductive load.
• Self-oscillating / hysteretic. Instead of a fixed external triangle, the loop oscillates at a frequency set by its own propagation delay and a hysteresis comparator. Inherently includes feedback (low distortion) and adapts its frequency to the signal, but the variable carrier complicates EMI filtering. Hypex UcD and Ncore designs are the famous high-end examples.

Failure modes — how Class-D amplifiers misbehave and die

• Shoot-through. If both MOSFETs in a half-bridge conduct simultaneously — even for nanoseconds — they short the supply to ground through a near-zero resistance. Current spikes to hundreds of amps and the devices fail explosively. Prevented by dead time, the deliberate gap where both gates are off. Too little dead time risks shoot-through; too much causes distortion.
• Dead-time distortion. During the dead-time gap the MOSFET body diode conducts the inductor current, and the exact output voltage becomes load-current dependent rather than purely duty-dependent. The result is a crossover-like nonlinearity that dominates THD at low levels. Adaptive dead-time control and post-filter feedback are the cures.
• EMI radiation. Fast edges (5–20 ns) at hundreds of kilohertz contain harmonic energy reaching tens of megahertz. The speaker cable becomes an antenna. Spread-spectrum clocking (dithering f_sw), good LC filter design, and slowing the switching edges (at a small efficiency cost) keep the amplifier under FCC/CISPR limits.
• Bus pumping. In a half-bridge driving low frequencies, energy is shoved back into the supply rail on alternate half-cycles, charging the rail capacitor and raising the rail voltage — which can trip overvoltage protection or stress the caps. Full-bridge BTL topologies and larger rail capacitance mitigate it.
• Thermal runaway in conduction. R_DS(on) of silicon MOSFETs rises with temperature, so a hot device conducts at higher loss, gets hotter still. Adequate copper, thermal pads, and over-temperature shutdown keep it bounded; this is far gentler than Class-AB thermal runaway because the steady-state loss is so much lower to begin with.
• Clipping and recovery. When the audio peak demands a duty cycle beyond 0% or 100%, the comparator saturates and the output clips. Recovery from clip must be clean or the loop "sticks" at the rail, producing a burst of full-power garbage. Anti-clip and modulation-index limiting circuits manage the boundary.

How Class-D compares to the linear classes

Property	Class-D (switching)	Class-AB (linear)	Class-A (linear)
Peak efficiency	85–95 %	~70–78 %	~25 % (50% theoretical)
Output device state	Fully on / fully off	Partly on (linear region)	Always conducting
Heat at full power (200 W)	~20 W	~100 W	~600 W
THD+N (modern)	0.005–0.05 %	0.002–0.02 %	0.001–0.01 %
Output filter	LC low-pass required	None	None
EMI risk	High (switching edges)	Low	Very low
Heatsink / size	Small	Large	Huge
Typical use	Phones, cars, PA, soundbars	Hi-fi receivers, guitar amps	Boutique hi-fi, headphone amps

The pattern is clear: Class-A buys the last fraction of a percent of linearity by burning enormous power; Class-AB is the long-standing compromise; Class-D trades a switching-EMI headache and an output filter for an order-of-magnitude reduction in heat. As feedback and process technology have closed the distortion gap, the only places linear amplifiers still dominate are where switching noise is unacceptable (sensitive measurement, some headphone amps) or where the designer values simplicity over efficiency.

Where Class-D shows up

• Smartphones and laptops. Mono or stereo 1–3 W filterless Class-D, often with an integrated boost converter that lifts the 3.7 V battery to ~9 V so the small speaker can get loud. Efficiency directly buys screen-on time.
• Soundbars, TVs, Bluetooth speakers. Stereo or multichannel 5–50 W Class-D, almost always full-bridge BTL with BD modulation. The TI TPA31xx and Infineon MERUS families dominate.
• Automotive audio. Trunk-mounted subwoofer amplifiers of 500 W to 2 kW are essentially only feasible as Class-D — a Class-AB equivalent could not be cooled in an enclosed vehicle. Modern OEM head-unit amplifiers are entirely Class-D.
• Professional touring PA. Powersoft, Lab.gruppen, and Crown rack amplifiers deliver tens of kilowatts from 2U chassis, with switch-mode power supplies feeding Class-D output stages. Weight savings are decisive for touring logistics.
• Active loudspeakers and studio monitors. Built-in Class-D amplifiers (often Hypex Ncore or ICEpower modules) let the speaker be self-powered with no external amp, with measured performance rivaling the best linear designs.
• Electric vehicles. Both the entertainment system and the legally mandated pedestrian-warning sound generator (AVAS) use Class-D for efficiency on the traction battery.

Common pitfalls when designing a Class-D amplifier

• Picking f_sw too high. Raising the carrier cleans up the PWM but multiplies switching and gate-drive loss linearly. There is a sweet spot (300–500 kHz for most audio parts); blindly going higher destroys the efficiency that justified Class-D.
• Ignoring the LC-filter–speaker interaction. An LC filter tuned into a nominal 8 Ω resistor will show several dB of response ripple into a real speaker. Either tune the filter for the actual load or use post-filter feedback.
• Under-budgeting dead time. Too little risks catastrophic shoot-through; too much dumps distortion at low levels. Adaptive dead-time control or a closed loop is mandatory for low THD.
• Treating EMI as an afterthought. Fast edges radiate. Spread-spectrum clocking, tight filter layout, and controlled slew rate must be designed in from the start, not bolted on after failing CISPR.
• Forgetting bus pumping. Half-bridge designs driving deep bass pump energy back into the rail. Size the rail capacitance and protection for it, or use a full-bridge.
• Neglecting idle loss in battery products. The constant gate-drive loss does not care how quiet the music is. In a phone, that standby current is the metric that matters most — design the modulator to gate the clock when silent.