APT0001
By: Richard Frey, P.E.
High Voltage, High Efficiency MOSFET RF
Amplifiers –
Design Procedure and Examples
1
High Voltage, High Efficiency MOSFET RF Amplifiers –
Design Procedure and Examples
Introduction
With the improvement in high power MOSFETs of late
– lower gate charge, low loss gate structures, and much
improved frequency capability – it has become more
possible to employ these “switchmode” devices in rf
generators at medium hf. The objective is to improve
the ‘power density’ (W/m3) and efficiency of the
equipment. The most common problem associated with
this application is getting the matching right. For the
gate drive side it is matter of obtaining an efficient and
repeatable match to a very small impedance. On the
drain side it is a matter of choosing a proper operating
class and power goal for the device and then designing
a network to provide repeatable and cost effective
realization.
The discussion below outlines the
procedure for both.
data sheet at the desired operating supply voltage. The
R is the effective series resistance, ESR, of Ciss and
includes dielectric losses,
gate metalization (or
polysilicon conductor), and bond wire conduction
losses. The L is essentially bond wire inductance,
generally package specific. The resonant frequency of
this RLC for large switchmode devices is usually in the
middle hf range, 10-30 MHz, and the Q is fairly high.
The Input
As an example, the APT 5010LLC data sheet Ciss is
5200pF at Vds of 25V. In practice, a higher drain
voltage is employed and this value will be somewhat
smaller. On data sheets where the C values are
graphed, it is common for the graph to stop at 50 or
100 volts. When graphed on a log-log scale, the C can
be extrapolated linearly to the proper operating voltage.
The effective gate C in an amplifier is increased
through the action of the Miller capacitance. Generally
the 50 or 25 volt value is a good place to start.
The gate of a power MOSFET can be modeled as a
series RLC circuit. The C is the specified Ciss from the
The gate ESR is seldom specified on the data sheet.
Measuring this on a RLC bridge using 5V of static bias
Figure 1 Input circuit design
2
with the drain-source shorted will yield a reasonable
value. The 5010LLC has a gate ESR of approximately
0.25Ω. This value is highly dependent on the die
geometry and fabrication process and varies widely
between manufacturers of otherwise similarly specified
parts. The higher this resistance is the lower the
frequency limit for efficient operation and the more
power will be dissipated in the gate structure. The gate
ESR of APT devices is very low because of the all
metal gate structure.
Gate parasitic inductance is generally between 5 and 15
nH depending on manufacturer, package and die size.
9 nH is a reasonable starting value for a TO-247 or TO264 plastic package.
The circuit topology chosen for the drive is a ferrite
loaded transformer with a single turn secondary. See
figure 1. The graphics for this figure were produced by
a Smith chart program called WinSmithi, which was the
tool used to design the networks described in this
article. The “Load R” is the gate ESR. It is not
represented in the series string in the figure but is the X
where the trace starts on the chart. The transformer’s
leakage reactance plus the gate inductance is series
resonated with the gate capacitance near the operating
frequency. Adjustment of the resonant frequency is
done using a series capacitor if the frequency is too low
– it usually is if the device is large. The gate is
referenced to ground or bias is supplied through a
resistor shunting gate to source. See the discussion on
biasing at the end of this article.
The procedure for designing the input circuit is fairly
straightforward. The circuit is constructed using
calculated and published values. Using an match
indicator connected to the drive input connector, the
circuit is energized with normal drain voltage, and the
circuit is adjusted for best match at the operating
frequency. The match indicator can be anything from a
network analyzer to a noise bridge. The requirements
are that it not inject enough signal to turn on the device
and it must be sensitive enough to indicate a match to
the driving impedance. The number of primary turns is
adjusted to bring the input impedance real part close to
50Ω and then resonate the resulting total inductance by
adjusting a series capacitor. In cases where a smaller
device is used and the input capacitance is insufficient
to resonate at the operating frequency, it may be
necessary to add a small inductance. If bias is going to
be used, this resonating should be done after the bypass
capacitor is installed because it adds ESR and changes
the total reactance.
The series capacitor is critical. The quality of the
capacitor determines the total ESR. Any loss in the
capacitor is drive lost and heating and resulting drift.
However, if the capacitor is large enough the loss may
be acceptable, or even desired if the gain of the
amplifier is already too high. Further, lossy capacitors
reduce the Q of the input and thus increase the
bandwidth and lower the tuning sensitivity. The
circulating current is high. Several capacitors in
parallel will provide adequate power dissipation if high
D caps are used and will offer a simple means of
adjusting the value. The voltage rating can be 100V.
For 13.5 MHz, the value used in the 5010LLC trial
amplifier was 5nF composed of five 1nF Z5U 500V
ceramic disc capacitors.
The transformer used is a two hole “binocular bead”
balun made from high permeability ferrite material.
The primary has ten turns of #28 insulated wire tightly
wound on the center of the core. Enamel coated wire is
suitable if care is taken to insulate the core. The
secondary is a single turn of 0.25” wide copper braid.
In this case, solder wick braid. This transformer may
run at considerable power. In the test amplifier, output
was 750W and the gain was 15dB, corresponding to an
input drive power of almost 25W. It runs warm. The
primary wire should have high temperature insulation,
enamel or teflon. The core material should have low
loss.
Ferrite bead balun cores of this type are available from
several manufacturers. Most are made for EMI
suppression and the core material is intended to be
lossy! In this case, the transformer core was from
Steward, part number 28N0765-100. Their type 28
material has an initial permeability, µi , of 850.
Steward’s type 25 material with µi =125 would have
less loss and almost equal leakage reactance. The
µi=850 material is more common and proved
satisfactory for the demonstration. However, in a
severe commercial environment, the µi =125 material
with its lower loss and higher Curie temperature would
be a better choice.
It is interesting to note that the driver is being matched
to the losses in the gate circuit. All the input power is
dissipated in these losses. What drives the MOSFET is
the voltage that appears across the input capacitance,
but no power is dissipated there. Also, because the gate
is a very low rf impedance, the signal that one can
measure at the gate terminal of the device is not the
signal across the gate capacitor. The gate inductance,
just 9 or 10 nH, is a significant part of the gate terminal
3
Figure 2 Output Network
input reactance. It can be calculated, but not measured
directly.
The Output
The load impedance required by an amplifier is
determined by the class of operation, the output power
and the supply voltage. The classic formula provided
in every solid state text book defines load resistance:
formula 1
RL= Vdd2/ 2Po
The factor 2 in the denominator works for linear
amplifiers running class AB where the drain swing is
reasonably sinusoidal and the conduction angle greater
than 180°. The factor diminishes with decreasing
conduction angle as the class of operation passes
through Class C to the high efficiency classes where
zero voltage switching is intended, like Class E. To get
the proper phase relationship between drain current and
voltage for these higher efficiency modes of operation,
the phase of the load impedance will typically be
positive.
The drain model of the MOSFET is the output
capacitance shunted by a “source resistance”. This is
necessary to perform the matching calculations but has
no physical realization. If there were such a source
resistance within the MOSFET drain, the efficiency
could never exceed 50%! The value of Coss can be read
from the data sheet. Its value should be that for half the
drain supply voltage to account for the nonlinear value
change over the total drain swing. The Q of the output
capacitance is a major contributor to efficiency. “Super
Junction” high voltage MOSFETs are notoriously poor
in this regard and should be avoided for rf amplifier
service.
The class of operation is defined by the drain
conduction angle; the portion of the operating
frequency period when the drain is conducting. Class
A operation is fully conducting, 360°. It never turns
off. It requires a positive dc bias voltage applied to the
gate. It is linear but not efficient. Class B is a special
case where the gate bias is adjusted for drain current
cutoff. The conduction angle is 180°. It is linear and
more efficient than class A but not more than 75%.
Class C is less well defined but is characterized by
conduction angles of less than 180°. There is usually
no bias voltage provided except by the drive signal, and
the efficiency is as much as 90%.
There are more classes. Saturated class C, sometimes
called C-E is a higher drive version of the same thing.
It can be 95% efficient. Class D is a particular
configuration using two transistors as switches. Its use
4
is limited by the parasitics of passive and active parts to
about 10MHz. It can be nearly 100% efficient. Class
E is a special zero voltage switching mode. It relies on
the proper phase relationship between drain voltage and
current to provide no overlap. The efficiency can be
100% with perfect parts. There are other classes,
somewhere between C and E, see US patent 5817580
for example.
For the purpose of this paper, the class will be sinewave
driven Class Eii. This does not require a large amount
close enough to permit minor tweaking for best
operation.
The output circuit is shown in figure 2. This is
deceptively simple and, with minor changes in part
values, drive level and bias, can be tuned to run any
class – A through Z. The calculation for load
impedance is per formula 1 except we will use a value
of 1.2 for the denominator factor. This compensates
for the large harmonic content in the drain waveform
and is partly the result of experience. The phase is
lagged somewhat to ‘unlap’ the voltage and current
waveforms and improve efficiency.
The APT5010LLC has a rated power
dissipation of 520W (only when the
junction temperature is at its
maximum of 150°C and the case is at
25°C). A practical useable value is
half that. This takes into account the
thermal resistance of a good drain
insulator and a lower maximum Tj. In
any good design, power dissipation is
determined
by
total
thermal
resistance.
Assuming
a
70%
efficiency for the amplifier, the
maximum output power can be
600W. If the actual efficiency is
better than this, the design margin
improves. 600W corresponds to a
drain load of 37.5Ω for a drain supply
of 150V.
Figure 3 Schematic as Built – 1kW @ 13.56 η=89%
of drive. At 13.56MHz, the small signal gain of these
MOSFETs is well more than 25dB. For saturated class
C, enough drive is used where it can be considered
almost a square wave equivalent. APT MOSFETs have
a gate breakdown voltage well past 40V. The proper
drain load has been determined empirically. In this
case, the load was measured after the amplifier had
been tuned for best gain and efficiency. The factor in
the load impedance formula was adjusted to fit the
result. The inductive component was derived from the
measured results and has proven effective in
subsequent amplifiers. In any case, the calculations
will yield the values needed for a functioning amplifier,
The output Coss is 1030pF at 25V –
not a particularly helpful figure. With
a 150V supply, the Coss is probably
closer to 700pF. This capacitor in
parallel with the 37.5Ω “source”
impedance will be used to determine the output
network. From this point the procedure is very
mechanical. The parallel combination is transformed to
an equivalent series impedance, 6.25 -j14. An L
network is used to transform this impedance to 50Ω.
How one chooses to design the L network is a matter of
preference. Handbook formulas will work as will a
Smith chart. The real part is transformed to 50Ω. The
required series L is increased by +j14 to cancel the
series output C.
In a practical circuit there are some slight adjustments
to this procedure. A method to feed the drain voltage is
Circuit diagrams external to APT products are included as a means of illustrating typical applications. Consequently, complete information
sufficient for construction purposes is not necessarily given. The information in this Application Note has been carefully checked and is believed
to be entirely reliable. However, no responsibility is assumed for inaccuracies. Furthermore, such information does not convey to the purchaser
of the semiconductor devices described any license under the patent rights of APT Inc. or others.
5
needed. A shunt choke is the typical method. If the
value is not too large, it will cancel some of the Coss. A
1 µH value was chosen which reduces the effective Coss
to 562pF. The calculation for the series equivalent
impedance and L network values is repeated. Some
form of blocking capacitor is needed to keep the supply
voltage out of the load. If placed right at the output
connector, the tuning capacitor will have both dc and rf
on it. It is thus better to put the capacitor in series with
the inductor unless the drain load is very low - as it
would be in a 50V rf device.
Choice of parts is a major factor determining the
overall success of the circuit. The passive parts are as
critical as the active device. As the stage power
increases, the stress on individual parts can be severe.
A case in point is the blocking capacitor. Placed in
series with the L network coil, the rf current through it
at 700W is considerable, 4.3A. Unless this is a low
loss capacitor, it will heat, drift, and even fail. Surface
mount parts are nice but not for this application. There
are “leaded surface mount” parts, stacked MLC types
made for SMPS service. The one used here is an AVX
part number SM067A822KAJ120. They are available
in 1kV COG characteristic from several sources in
values up to 10nF. A high Q coil is necessary. Placing
this coil near a ground plane or compartment wall will
lower the Q and increase the losses. It can get hot
enough to unsolder itself!
Figure 4 5010LLC Drain voltage waveform
Tuning the output circuit is a matter of adjusting the
series coil L2 and shunt capacitor C4 for best output
power and efficiency. Initial tuning is best done at a
low duty cycle, 20 to 50 %. For a trial value of L2 ,
adjust the value of C4 for best performance. Adjust L2
slightly and repeat the sweep of C4. Make sure the
most efficient power does not result in a peak voltage
swing too close to the BVdss rating for the device. In it
not unusual for the most efficient power to be well in
excess of the voltage capability when the conduction
angle and drain current-voltage overlap are minimized.
If the peak voltage is too much, add some external Coss,
reduce L2 and tune for a lower power, or reduce the
supply voltage. The peak voltage can be three to four
times the supply voltage.
The circuit of Figure 3 was also used with an
APT10050LVR, a 0.5 Ω 1kV part similar to the
5010LVR. With only slight adjustment of L3 and C2,
similar results were obtained.
Note in both Figure 4 and Figure 6 that the peak drain
voltage is more than the rated breakdown voltage of the
part used. Clearly, there is some avalanche occurring
but the total avalanche energy is very small as indicated
by the good efficiency. This is not a desirable situation
and can be remedied in several ways. Reducing the
supply voltage is the most obvious but is not always an
option. The next best way is to reduce the impedance
of the drain circuit by adding additional shunt
capacitance across the drain and retuning the output
network. In so doing, the peak drain voltage is reduced
without materially effecting the output power because
the impedance is also lowered.
There are programs available which permit designing
analyzing, and optimizing class E performance before
doing anything in the lab. This can be very useful for
evaluating and comparing different devices, and for
adjusting a circuit for a particular maximum peak drain
voltage.iii
It will be necessary to go back and readjust the input
match because Miller capacitance, Crss, does indeed
change the input impedance. Watch for signs of
instability – the tendency for oscillation. If the layout
is poor, or the gain is too high, it will oscillate. There
is another phenomenon called snap-on which occurs
when the change in input impedance between full drive
and no drive interacts with the driver. When the input
has been adjusted for best SWR at full output the driver
provides maximum power to the gate. When drive is
reduced, the output goes to zero. When the drive is
increased from zero, the device input impedance is
different and the driver will not be able to provide
enough drive at the normal setting. As it is increased
further, still with a fairly poor match, the input will
eventually have enough drive to make output. As the
Crss feedback changes the input match, it improves, the
power transfer into the gate improves and suddenly we
have full output.
6
Figure 6
This can be seen as drive hysteresis. A stiffer driver,
more series resistance in the gate, or additional fixed
shunt capacitance across the gate will improve this
situation. In the test amplifier of Figure 5, 1nF was
added across each gate to ground. The input match was
readjusted and the hysteresis was almost completely
eliminated.
If several devices are being paralleled, it is necessary to
insure they will share the load.
It a common
misconception that paralleled MOSFETs inherently
share the load. No, they do not!
Even
though
the
temperature
coefficient of RDS(ON) is positive, this
only effects the fully saturated
condition.
The
temperature
coefficient of the gate threshold
voltage is negative. With constant
voltage bias, the drain current goes up
as the device heats up. The hotter
device is turned on harder.
In an ideal device that is conducting
current and dissipating power, it can
be assumed that current distribution
across the die is uniform and that the
distributed junction temperature is
also uniform. In reality, however, if
there is any trace of non uniformity in
the current distribution, the junction
temperature will be higher in those
regions where the current is highest.
Non uniformity can be precipitated by
small voids in the die
attachment
solder,
non
uniform junction doping, and
by a host of other variables.
Hot spotting is then triggered
by the negative temperature
coefficient of gate threshold
voltage acting in a localized
manner. As one part of the
die heats up more than
another, the local junction
Vgs(th) falls and Gm then
causes that area to turn on
harder drawing even more
current. Since Gm increases
with increasing current, the
runaway process is accelerated
even more. The fact that
has
a
positive
RDS(ON)
temperature coefficient is
irrelevant in that the MOSFET
is operating in the linear
region, and not in "saturation".
The bigger the MOSFET die, the higher the power
dissipation and the greater probability of experiencing
this problem. The alternative is to use multiple smaller
devices and buffer them with source degeneration
resistors.
Gate Bias
Figure 5 2x ARF448 waveforms. Po = 600W, η = 90%.
7
In an N-channel MOSFET, zero volts between the gate
and source results in no drain current; the device is cut
off. These devices can run class C without any
additional voltage on the gate. To run class B, a
positive voltage is applied to the gate to a point where
the device just starts to turn on. Class AB bias would
be set for some amount of static no signal drain current.
The gate threshold voltage of all power MOSFETs has
a negative temperature coefficient. It is usually shown
in a figure on the data sheet as Id vs Vth at several
temperatures. It means that a fixed amount of class B
bias applied to a cold device can be enough to put it
well into class A after the application of rf drive due to
the change in Tj. This will kill the part by over
dissipation if the drain supply voltage is more than 75V
unless a system is in place to actively adjust bias for
changes in Tj.
High voltage MOSFETs do not work well with forward
bias but bias adjustment can be useful. It has been
mentioned in the literature for years that one of the
benefits of the MOSFET is the ability to control the
gain of the device by adjusting the gate bias. The
typical gate threshold voltage is 4V. That means an 8V
p-p signal is needed on the gate to start turning it on. If
1V of fixed dc bias were put on the gate, the effective
threshold would be reduced by that amount and a 6V pp signal would have the same effect on the drain. The
gain is higher. The reverse is true. A negative bias will
increase the threshold and effectively reduce the gain.
This has great implications for gain/power control.
Systems
The output power required for a typical generator
system is always more than can be obtained from a
single device. In order to minimize the amount and
cost of the “glue” required to drive and combine the
outputs of several amplifiers needed to produce the
required power, the opportunity exists to use custom
devices containing multiple die. The motivation is
improved reliability, efficiency, power density, and
lower cost. Very large die are subject to hot spotting.
Several smaller die with suitable buffering can be
incorporated within a thermally efficient package to
improve the performance and increase ruggedness and
reliability. The cost of the package is more than offset
by the savings in combining parts and circuit
simplification.
to guarantee stability under all load conditions. Class
C-E or saturated C and class E amplifiers have less
tolerance to load disturbance. In order to provide a
working system in 13.56 MHz plasma generators, a
matching device is needed between the load and the
amplifier. It will ‘follow’ load changes and provide a
more constant load to the amplifier.
Harmonics are not a real problem for the plasma
generator. While some specifications put limits on
harmonic energy delivered to the load, the effectiveness
of the plasma itself is not effected by harmonic energy.
The most significant problem is the effect of the
harmonic energy on the VSWR detectors used to
control the matching network. It is sometimes useful to
place a lowpass filter on the rf samples before detection
to insure a good null at the operating frequency. If an
output low pass filter is used, its effect on the stability
of the amplifier must be considered. Harmonic energy
reflected from the lowpass filter circulating in the
output network can cause serious component heating,
loss of efficiency through modification of the drain
waveform, and instability when the match is disturbed.
Use of a highpass/lowpass diplexeriv with the highpass
section terminated in a load can ease the problems
associated with harmonic energy.
Conclusion
If proper drain load is chosen, the output power and
efficiency goals for the amplifier stage can be reached.
Proper functioning of the circuit in a commercial
environment is very much dependent on the quality and
selection of the passive parts in the circuit as well as the
cooling and load protection systems necessary to
support it.
i
winSMITH 2.0, 1998, available through Noble
Publishing Corp.,4772 Stone Drive, Tucker, GA
30084.
ii
J. F. Davis, D. B. Rutledge, “Industrial Class-E power
amplifiers with low-cost power MOSFETs and sinewave drive” Conference Papers for RF Design ’97,
Santa Clara, CA, Sept 1997, pp. 283-297.
iii
N. Sokal, HEPA-PLUS/WB v3.29, Design
Automation, Inc., Lexington, MA
iv
Sabin, Schoenike, et al, Single Sideband Systems and
Circuits, McGraw-Hill 1993, p. 492.
The more efficient the amplifier, the more dependent it
becomes on having a particular load for proper
operation. Class A amplifiers are most often used in
wideband general use lab amplifiers which are required
8