New approaches to switched-

mode audio power amplifiers

Tom Lawson -March 06, 2013

The Class D amplifier has been the usual switched-mode answer to linear amplifiers for
three decades. Class D amps have a clear efficiency advantage over their linear
alternatives, which, over the years, has led to much effort being invested in improving the
linearity of switched-mode designs.
Even so, the harmonic distortion of switched-mode amps remains inferior to that of linear
power amplifiers. A new approach promises better efficiency than Class D amplifiers and
better performance than linear amplifiers.
The Basic Class D Amplifier
The basic Class D audio amplifier is shown in Figure 1. Either the upper or lower switch
in the totem pole is ON at any time, except for a brief transitional period which prevents
both switches conducting at the same time. The two switches operate alternately to
produce an average output that remains between the Power Input voltage and ground.
The time ratio of the ON time of the two switches determines the average output voltage.
For example, a 50% duty cycle will produce an output of half the Power Input voltage.
The series capacitor removes the DC component. The series inductor and shunt
capacitor form an LC low pass filter. The switches operate at a frequency at least least
ten times higher than the maximum output frequency to allow the filter to remove the
chopping frequency from the desired output signal.

Figure 1: The basic Class D switched mode audio amplifier shown here in block
form
A simple amplifier of this sort is entirely adequate for voice applications, but there are
implementation details that need careful attention for more demanding uses. A cell
phone audio amplifier is typically a Class D amp. These amplifiers were originally thought
of as needing only to meet the standards for voice communication. Increasingly, phones
and handheld devices are being called on to produce high fidelity audio. With care to
minimize distortion, Class D amplifiers can still meet that need, though their limitations
are too often plainly audible.
There are three main sources of distortion in Class D Amps. First, the best FET switches,
driven optimally, still spend a significant time in the linear region when switching. That
means dead time must be inserted to avoid upper and lower FETs conducting at the
same time (shoot-through). Dead time causes non-linearity in the output which is difficult
to correct for. The brief period of linear switch operation is a source of asymmetry which
contributes to distortion, as well as reducing efficiency.
Second, Class D amplifiers do not provide power supply rejection. Audio frequency noise
on the supplies will appear only slightly attenuated at the output. That requires extra care
in regulating and filtering the power supply. The task is further complicated by the way
Class D amps return inductive energy to the rails, sometimes called bus pumping. Bus
pumping can cause a power rail to rise above its regulation point, leading to distortion.
Third, the switched inductor in a Class D amp interacts with the speaker's inductance (or
capacitance) in non-linear fashion. Care must be taken to avoid resonances and beat
frequencies that can be heard in the audio range.
Faster clocking decreases the size of filter elements and extends the upper range of the
bandwidth, but it also increases switching losses. As the clock rate increases, dead time
becomes a higher percentage of the cycle time, which increases non-linearity. This effect
can quickly become the largest source of distortion.
Take the example of a Class D amplifier running at 1 MHz with 25 ns of dead time. At a
50% duty cycle, after subtracting for the dead time, the upper switch is on for 475 ns and
the lower switch is on for the same 475 ns. The theoretic output is then exactly where it
should be, at 50% of full scale. The situation is different at 25% of full scale. On time is
now 175 ns and OFF time is 775 ns instead of 200 ns and 800 ns. That puts the theoretic
output at 175 / 775, or 22.58%. There is almost 2.5% distortion before the non-idealities
of the power switching are taken into account.
While it is possible to reduce dead time below 25 ns, extreme care must be exercised.
Temperature and aging effects can shorten dead time, causing shoot-through. Even a
little shoot-through disturbs the power rails and generates distortion. A little more shoot-
through can be destructive.
A related issue is the difficulty in producing extremes of duty cycle that still exhibit
symmetrical positive and negative edges, and that do not dwell too long in the linear
switching region. The problem stems largely from the limitations of switch drivers.
One answer is to use Sigma-Delta modulation instead of Pulse Width Modulation. With
Sigma-Delta modulation, periods of ON or OFF are integer multiples of the clock period.
Linearity can then be improved, but at the expense of a higher clock rate. Sigma-Delta
systems generally need to run at least 64 times the maximum signal frequency. There
are significant efficiency penalties associated with clocking that much faster.

Even after taking careful measures to minimize the error sources cited above, negative
feedback is required to obtain good performance. Negative feedback is problematic in a
Class D amplifier because the output filter necessarily inserts a delay, which tends to
destabilize any feedback loop.
The Class D amp is essentially a buck converter. Buck converters need compensation to
remain stable for the same reasons. In order for feedback from the output to be applied
without inducing a tendency toward oscillation, the feedback must be processed. That
processing must be done with care to avoid causing frequency-dependent delays
appearing as distortion. Otherwise, the feedback must be taken before the filtration,
which means the non-linearity of the filter, and the interaction of filter and load reactance,
is outside the loop.
The conflicting aims of faster correction of errors and limiting the tendency toward
unwanted oscillation are all too familiar to those who work with this topology. Ultimately,
it is this compromise which defines the limits of performance for Class D amplifiers.
Another limitation of the Class D amplifier of Figure 1 is easier to address, i.e., the limited
output voltage swing. In a cell phone, the power input is apt to be 4 volts, and a 4-volt
output swing is often not enough for audio. The solution is the Bridge Tied Load (BTL)
topology. In Figure 2, below, a Class D BTL amplifier is shown driving a piezo speaker,
which is essentially capacitive. Piezo, or ceramic, speakers are desirable in cell phones
because of their thinness and efficiency, but they require higher voltages to drive them
properly.
The Bridge Tied Load Variation on Class D amplifiers

Figure 2: Here, a Class D BTL amplifier drives a piezo speaker.
The differential voltage across the speaker can now be twice the Power Input voltage. In
operation, both totem poles need one, and only one, switch to be ON at any time.
The efficiency of the BTL amp can be very good, but it is reduced by the need for a
second switching element in series with the power path. The BTL form is more flexible,
and it exhibits less bus pumping. In the piezo speaker case, a larger efficiency issue
appears. Series resistance needs to be added to prevent oscillations caused by the
capacitive loading. That resistance can easily become the largest single source of
inefficiency in the system. On balance, the compromises work out to favor the BTL
configuration for handheld devices.

Filterless Variations
In some cases, the reactance of the speaker itself is relied on to filter out the chopping
frequency. That can be satisfactory, but it requires the speaker to dissipate extra power
due to the high-frequency ripple current. Not all dynamic speakers are sufficiently
inductive at those frequencies to be efficient as a filter. Equivalent Series Resistance,
ESR, causes analogous losses in piezo speakers.
In either case, efficiency suffers, the speaker heats, and the sound quality is
compromised. Also, if the speaker is at any distance from the amplifier, the wiring is
required to carry square waves. The edges are attenuated by stray inductance,
degrading the sound, and those edges are a source of electromagnetic interference.
Care must be taken to avoid damage to dynamic speakers in the filterless case. Above
the audible range, voice coil movement is proportional to 1 / frequency
2
. If the modulation
frequency is high enough, the movement is small, so the voice coil won't hit the limit of
travel and cause damage. As long as the speaker is sized to dissipate the extra energy,
the filterless variation can be satisfactory. An additional efficiency implication is
concealed here. You cannot simply reduce the modulation frequency to save power,
without putting the speaker at risk.
Filterless amplifiers for piezo speakers do not need to take voice coil damage into
account. These amplifier/speaker combinations are the most efficient option now
available. One manufacturer documents power consumption approaching half the power
of Class D driving a dynamic speaker. That comparison includes a 10 ohm series resistor
for stability in the piezo case. The largest single source of inefficiency is that resistor, so
there is still room here for improvement.
The awkward overall conclusion is that the efficiency advantages of Class D amplifiers
stem from the digital nature of their output, while the performance disadvantages of
Class D amplifiers stem from that same digital nature. What is really called for is a digital
amplifier with an analog output. There is now an option in that category, enabled by
improved FET switches and a control strategy called Predictive Energy Balancing (PEB).




Principles of Predictive Energy Balancing
Predictive Energy Balancing, (PEB), allows immediate, direct feedback in a switched-
mode power converter by effectively removing the filter pole from the feedback loop.
Predictive Energy Balancing makes switching decisions based on the eventual outcome
of those decisions, which can be calculated in advance using the fundamental energy
equations. In the process, the unavoidable delay of the output filter is removed from the
feedback loop. PEB allows the maximization of both agility and stability.
The mathematics for energy balancing in any type of power converter can be derived
from a few fundamental formulas. The kinetic energy held in an inductor, L, is:
KEL = (I
2
x L) / 2
where KEL is inductive energy in joules, I is current in amps, and L is switched
inductance in henries.
The kinetic energy held in a capacitor, C, is:
KEC = (V
2
x C) / 2
where KEC is capacitive energy in joules, V is voltage in volts, and C is filter capacitance
in farads.
In general, the inductive energy term represents the Supply and the capacitive energy
term represents the Demand. The Demand is the difference between the instantaneous
capacitive energy and the desired capacitive energy. At the regulation voltage, Reg, the
energy held in the output filter capacitor would be:
KEReg = (Reg
2
x C) / 2
where KEReg is desired capacitive energy in joules, Reg is Reference point in volts, and
C is filter capacitance in farads.
The capacitive energy deficit is then KEReg - KEC.
For a simple, discontinuous flyback converter, the energy balance point is the moment in
time when the inductive energy is equal to the capacitive energy deficit:
KEL = KEReg - KEC
In an energy balancing power converter, during a control cycle, the supply term, KEL,
rises exponentially while the demand term, KEReg - KEC, rises in piecewise linear
fashion. The balance point is reached when the Supply equals the Demand. Then,
inductive energizing stops, and the inductive energy is transferred from the switched
inductor to the output filter capacitor. With the balance properly scaled, after the inductive
energy has been transferred, the voltage on the filter capacitor will equal the regulation
voltage.

Figure 3: Simulation showing the operation of PEB controls for a switched mode
amplifier
The figure above illustrates a PEB amplifier in operation. A simulation is used here for
clarity, though the waveforms on the bench are substantially the same. Each control
cycle here is 5 s long. Demand has been rising since the end of the previous Transfer
period. The Energize period begins with the Supply term at zero. Supply rises
exponentially until the balance point is reached, which causes switching to begin the
Transfer period. Transfer continues until the inductive energy has been moved to the
output. At this point, both Supply and Demand are near zero, and the Output closely
matches the Reference voltage. An Idle period follows, waiting for the next control cycle
to begin.
Notice that the Idle period can be short or long. The PEB calculation is not affected by
the conversion rate. Also note that each control cycle is self-contained. There is no
averaging or history involved. That property allows instantaneous feedback without
inducing a tendency toward sub-harmonic behavior.
















The Predictive Energy Balancing Amplifier
Because PEB allows an output to be regulated precisely on a cycle-by-cycle basis,
instead of on the average, PEB enables an entirely new form of switched-mode amplifier.
The following block diagram shows one version of this new class of amplifier. At first
glance it appears similar to the H-bridge configuration of a Class D amplifier, but with
some of the components rearranged, and with different controls. In operation, it is quite
different, being essentially a bidirectional flyback converter. Figure 4 shows one form of
PEB amplifier driving a piezo speaker, though the same amplifier will drive a dynamic
speaker, as well.

Figure 4: A Block Diagram of One Form of PEB Amplifier
The PEB calculations are done in real time by the circuitry on the left side of the block
diagram. These calculations can be done in analog or digital fashion. We have chosen
simple, low power analog circuitry for the demonstration systems. The squaring circuits
respond in a few nanoseconds and consume only micro amps of current. A reasonably
fast differential amp and comparators will suffice. Synchronous rectification is used for
efficiency in the demonstration system, but is not essential.
There are no requirements here that stretch the limits of conventional practice.
Eventually, these calculations should be integrated for cost and size advantages. The
circuitry on the right in Figure 4 consists of two ordinary totem poles forming an H-bridge.
BTL Class D audio amplifiers would have the speaker in place of L1. Here, the speaker
is placed between the output and the power supply. Depending on the process and
power level, the power circuitry can be incorporated into the same integrated circuit.
In operation, the amplifier uses flyback energy transfers to push the output away from
zero. It uses reverse flyback transfers to draw energy back from the speaker to pull the
output towards zero. The inductor in a PEB amplifier is active at the switching frequency,
so a much smaller inductor filters the digital power switching to produce an analog
output, when compared to the inductors needed to filter Class D amplifiers. PEB
amplifiers can be fully bipolar, or can be offset to drive the output above and below the
power supply voltage, as shown here. The speaker could just as well be connected
between the output and ground with a coupling capacitor to remove the DC bias voltage.


Figure 5: Screen shot of a PEB amplifier producing a 10 kHz sine wave 8 volts
peak to peak driving a piezo speaker
The 200-kHz modulation frequency can be plainly seen in the output trace. A minimal low
pass filter will remove the remnants of the modulation, if desired. The reference trace
represents the ideal output. The narrow peaks in the output trace, every 5 s, indicate
the regulation points. Those peaks are caused by the ESR of the speaker's capacitance.
The inductor current flows out of the battery when the voltage at the output increases,
and flows back into the battery when the output voltage decreases. At lower frequencies,
or lower amplitudes, less energy needs to be moved to keep the loop in balance.
Because feedback acts to control regulation on a cycle by cycle basis, the difficulties of
Class D feedback are avoided. Any changes related to power supply voltage, dead time,
or the impedance of the load are corrected inside the loop in the course of each control
cycle.







Cautions
Because of their differences from Class D amplifiers, there are a few cautions to observe
when applying PEB amplifiers. The gain of an amplifier refers to the amplitude of the
output signal relative to the input signal, but there is second gain term in a PEB amplifier
that governs its dynamic behavior.
Because the energy balance scaling is in proportion to the ratio of the switched
inductance and the load capacitance, the PEB gain needs to be matched to the
capacitance of the load. One consequence is that inductors showing large changes of
inductance at different currents should not be used for PEB applications to avoid causing
non-linearity.
In the case of a dynamic speaker, which is inductive, not capacitive, a capacitor is added
at the amplifier output to set the PEB scaling and the output ripple. Then, the resistance
or inductance of the speaker has little effect on the dynamic performance. For piezo
speakers, the PEB scaling must be set to match the chosen speaker's capacitance.
That limitation is not inconvenient in the case of a handheld device with built-in speakers,
or in the case of a powered speaker, but it could be an issue for a general purpose
amplifier. If the ability to dynamically adjust for varying capacitance at the load is needed,
an additional circuit must be added. Conventional negative feedback, heavily filtered to
have only a DC effect, can be added to set the PEB scaling.

Efficiency comparison of PEB and Class D amplifiers
The theoretic maximum efficiency of Class D and PEB amplifiers is 100%. As a practical
matter, both types of amplifiers can be expected to run in the region of 90% efficiency.
There are five factors which can give the PEB amplifier an incremental efficiency
advantage over Class D amplifiers:
1. First, Class D amps need to be clocked faster because of the digital nature of
their outputs. A PEB amplifier can run at half, or one-fourth, the clock frequency
and still offer superior fidelity.
2. Second, the clock frequency does not enter into the PEB equations, so the
dynamics remain intact as the clock frequency changes. That enables dynamic
clocking to be used to save power with little effect on fidelity.
3. Third, the greatly improved power supply rejection of PEB amplifiers eliminates
the losses associated with decoupling and re-regulating power supplies.
4. Fourth, almost all the losses due to ripple current in the speaker are eliminated in
the PEB amplifier.
5. The last factor applies only to piezo speakers. Class D amplifiers need series
resistance added to remain stable when driving capacitive loads. A typical
minimum value for that resistance is 8 ohms. The resistor is dissipative, and can
account for the majority of power losses in a Class D amp, greatly reducing the
efficiency advantage of piezo speakers. PEB amplifiers do not need such series
resistance.


Conclusions
By rearranging the same circuit elements now employed in BTL Class D amplifiers and
by adding predictive controls, real improvements in fidelity and efficiency of power
amplifiers have been achieved. Since changing the control intelligence does not add to
the cost of an integrated circuit, these benefits can be realized without an increase in
cost. In fact, system costs can be reduced by eliminating filter inductors and by
generating less waste heat. PEB amplifiers are particularly well suited for driving piezo
loads.
The same audio amplifier technology can be scaled for hearing aids, or consumer
electronics, all the way up to theater-sized systems. Related versions of this technology
can be adapted to motor drives and other power amplifier applications, as well as to
power converters, in general. The PEB amplifiers shown here are covered by four issued
US patents. Other patents are pending.

About the author
Tom Lawson began working in instrumentation in 1968. He formed Lawson Labs in
1981, and CogniPower in 2009. His patents span five decades.