Professional Documents
Culture Documents
Abstract
by
Anthony John Stratakos
CMOS design have assumed that the supply voltage is a free variable and can be set to
any arbitrarily low level with little penalty. This thesis introduces and demonstrates an
array of DC-DC converter design techniques which make this assumption more viable.
converters are summarized. Design techniques at the power delivery system, individual
control system, and circuit levels are described which help meet the stringent
expressions for losses are presented. Special design considerations for the key dynamic
voltage scaling enabler, called the dynamic DC-DC converter are given. The focus
throughout is on low-power portable applications, where small size, low cost, and high
The design and measured results are reported on three prototype DC-DC
converters which successfully demonstrate the design techniques of this thesis and the
processor. This is made possible by a new class of voltage converter, called the dynamic
DC-DC converter, whose primary performance objectives and design considerations are
Table of Contents
Chapter 1: Introduction ....................................................................................................1
1.1 Motivation......................................................................................................................1
1.2 The Challenge of Lower-Voltage DC-DC Conversion ..................................................3
1.2.1 Low-Voltage and High-Current.................................................................................................4
1.2.2 Low-Voltage and Low-Current .................................................................................................5
1.3 Research Goals and Contributions.................................................................................7
1.4 Thesis Organization .......................................................................................................8
References.......................................................................................................................231
Acknowledgments vii
Acknowledgments
It has been an honor and a privilege to study at Berkeley. There are many
people to thank: Those who inspired me, those who provided technical guidance, and
those whose friendship made even the most difficult times more enjoyable. Most of the
people I list below have provided inspiration, guidance, and friendship. To these
Before anyone else, I must thank my brother. From early childhood to today, I
have excelled mainly by following his example. I will always admire him and he will
From both Mom and Dad, I learned methodical and analytical thought. Sorry Dad: Any
Jolie Kerns continued to feed my creative side and has offered the
encouragement to make it through the last three years. I’m not sure I would have made
it without her. With my own parents 3000 miles away, Trish and Gary have provided a
From our very first 140 problem set through our theses, Dave Lidsky and I
have been partners and best friends. We grew together, but perhaps I more than he: He
taught me to find the essence of a design, a talk, or a paper, and showed me how to get
the most out of grad school − by learning a little bit from every person around me. We
also had a lot of good times; I think my right arm is six inches longer than the left from
throwing every imaginable type of spherical object at Dave and his pet rodent,
Satchnomo. (Yes, Eleta, that was Dave’s dog you smelled every night and weekend and
summer day.) And you meant “back” rather than “backside”, right?
Acknowledgments viii
Andy Abo was my housemate and close friend for six years. We endured a lot
of school-induced pain together, but always survived. I thank him for his friendship,
good humor, and turkey tacos. Andy also taught me an important trait: Moderation. But
has he really never seen a great movie? My parting advice to him: Stay off the court!
As Dave and Andy helped me to grow, Chris Rudell did all he could to stunt
that growth. With four years as housemates, and countless trips to the RSF and Tilden
Park, I can guarantee that I have heard every one of his hilarious stories and seen every
we hadn’t (wasted? ... I don’t think so) so much time watching sports together. Sekhar
shared good music, good books, videotapes, and many laughs. He is underrated as a
Jeff Weldon and I shared trips to the RSF, Arinell’s, Lo Cocco; NBA; the city.
He’s the one with whom I thoroughly appreciated the finer points of going to school in
the Bay Area. He also served as my fashion consultant. Thanks Jeff, I will never wear a
exceptional people. I learned something from each of the following: Arthur Abnous,
Arya Behzad, Paul Haskell, Srenik Mehta, Keith Onondera, Craig Teuscher, Marco
Zuniga. We will be friends forever. Special thanks to the 920 Keeler founders, Srenik
Mehta and Arya Behzad, and to Katerina Pappas for renting us such an awesome house.
I am grateful to Rhett Davis for helping with the design of the DVS chip.
I learned a lot by working with Tom Burd and Anantha Chandrakasan. They
provided the low-power applications which drove my research and gave me lots of good
advice. I’ll also remember Tom’s bachelor party forever. It is either that or Chris’ 30th
birthday party which I rank as the single best night of my grad school career.
Acknowledgments ix
Andy Burstein, Cormac Conroy, Greg Uehara, and Sam Sheng were critical to
Brian Acker and Charlie Sullivan were colleagues and friends. Brian helped
me in the lab, inspired and validated a great deal of my work, and showed me some
good mountain bike trails in Tilden Park. Charlie was a mentor. I could go to him with
any problem, technical or other, and come away with the answers I needed. We also had
Bob Brodersen provided creative advice, research focus, and first class
facilities (Did your advisor rent the Monterey Bay Aquarium for a research retreat
provided guidance and strong technical support. While I learned a considerable amount
about power circuit, control system, and analog IC design from Seth, I may be most
grateful for his contributions to my writing style. Thanks for making me feel like an
I am especially grateful to Jan Rabaey and Bob Meyer for teaching me so much
about digital and analog circuits. Through their instruction, I learned to think
intuitively about circuits, and learned how to pursue research. Because they both have
strong personalities and good senses of humor, their lectures were usually a lot of fun,
too.
Tom Boot, Heather Brown, Peggye Brown, Ruth Gjerde, Elise Mills, Carol
Sitea and Kevin Zimmerman made sense of the confusion that is UC Berkeley. I feel
like they all went out of their way to help me at various times.
1.1 Motivation 1
Chapter 1
Introduction
1.1 Motivation
supplies. Portable electronic equipment, such as laptop computers and cellular phones,
such systems is to run the digital CMOS circuits at the lowest possible supply voltage
and compensate for the resulting decrease in performance with architectural, logic-
run-time, dynamically adjusting the supply voltage to trade performance for energy
power design strategy assumes that the supply voltage is a free variable and can be set
to any arbitrarily low level with little penalty. In portable electronic systems, high-
terminal, including the digital baseband circuitry, and speech, pen, and text/graphics I/
(from 1.5 V to 1.1 V), low-current (as low as 5 mA) DC power supplies must be
supported by a single battery source. Because the system also requires supplies of +/- 5
V and 8 V to power the flat panel display, RF transceiver circuitry, and microprocessor
subsystem, a total of six voltage converters are needed to generate all of the voltages
from a single 9 V battery source. These converters consume 42% of the overall power
and 12% of the system volume of the Infopad [Truman98], and cost as much as 54
dollars 1 .
Voltage regulation as an interface between the battery source and load can
further enhance battery run-time. A circuit may be designed such that its optimum
operating voltage is the end-of-life voltage of a specific cell, apparently minimizing its
power consumption without the use of a DC-DC converter. This not only makes the
circuit design challenging (the voltage of a typical AA-type lithium ion cell may vary
by as much as +/- 20% of its nominal value throughout its discharge), but because the
cell discharge characteristic is not flat, the circuit will consume greater than its
minimum operating power from the cell throughout the majority of its discharge. If a
DC-DC converter is inserted between the cell and the load, and the converter’s output
voltage is maintained down to the end-of-life cell voltage, the circuit will consume its
system run-time (by as much as 50% for a digital CMOS circuit powered by a single
1. Cost estimate based on IC and all external components purchased through a distributor in 1000 quantity.
1.2 The Challenge of Lower-Voltage DC-DC Conversion 3
battery run-time. Power management schemes are used in most low-power hardware:
Unused circuitry is powered-down and gated clocks are employed to reduce power
techniques may present severe load variations (up to several orders of magnitude), and
the system may idle for a large fraction of the overall run-time. This implies the need
for a high conversion efficiency not only under full load, but over a large load
systems, the quiescent operating power (control power) of the regulator must be kept to
an even lower level to ensure that it does not contribute significantly to the overall
[Chandrakasan94a] which supports speech I/O, pen input and full motion video, and
consumes less than 5 mW at 1.1 V. The control circuit for a converter supplying this
mass. While high-efficiency DC-DC conversion can substantially improve system run-
may also be achieved by simply increasing the capacity of the battery source. However,
custom-designed to their individual loads, their volume will typically be much smaller
than the volume of the additional battery capacity required to achieve the equivalent
extension of run-time.
DC-DC conversion, each with a unique set of challenges: Low-voltage and high-
1.2 The Challenge of Lower-Voltage DC-DC Conversion 4
current; and low-voltage and low-current. While both are summarized below, this thesis
devices where high efficiency is crucial to maximize battery run-time, and small
Although voltages continue to scale downward, rapidly approaching 1.5 V and below,
both clock speed and physical capacitance increase with decreasing feature size,
converter supplying this current had an effective series resistance of only 10 mΩ due to
the sum of the on-resistance of the FETs, all series resistance associated with bonding
and packaging, and the equivalent series resistance of the filter inductor and its
interconnection, the converter would be only 60% efficient − before all other losses
were considered. The resistance from ten squares of standard one ounce printed circuit
board copper would alone contribute nearly 25% loss. Such problems are unlikely to be
solved with clever circuit design. New parallel power supply architectures, flip-chip
solder bump and micro-BGA assembly technologies, and chip- and board-level
2. Based on scaled Pentium Pro current and voltage demands of 13 A at 2.4 V [Intel97]. Assumes process
technology scaled to 0.18µm with appropriate voltage scaling and an increase in average chip power consis-
tent with technology scaling trends [Rabaey96].
1.2 The Challenge of Lower-Voltage DC-DC Conversion 5
Worse still, is the rate at which such a processor demands its current. It can
transition from sleep mode to full operation in a time scale of nanoseconds, presenting
a load step as high as 40 A to the output of the DC-DC converter. This transient requires
a huge amount of bypass capacitance to maintain a stable voltage at the processor pins.
for adequate bypass decoupling [Arbetter98]. With the higher current demand and
tighter voltage tolerance of next generation processors, this capacitance seems destined
to exceed 10 mF, with an ESR requirement of less than 1 mΩ. This problem is currently
being addressed at the circuit-level, with the introduction of the active clamp [Wu97]
presented by specialty digital signal processing ASICs for portable electronic devices.
Here, the digital IC is typically designed to meet a certain throughput constraint, often
to the voltage scaling techniques presented in Section 2.1, and unlike a general purpose
processor, its current consumption scales with its voltage supply, resulting in lower
power seems insignificant in nearly any real-world application, in many cases, it is not.
Consider a cellular phone or pager in standby mode. While the higher-power RF and IF
receiver components are pulsed with a small duty cycle, a variety of specialty and
power digital hardware is often the limiting factor in standby battery run-time.
1.2 The Challenge of Lower-Voltage DC-DC Conversion 6
A DC-DC converter supplying such a load must, itself, be far lower power than
that load. This presents a number of challenges, many of which require circuit
innovation. For example, a 1 MHz PWM converter powered by a single lithium ion cell
capacitance. It is feasible that the connection of the external filter inductor alone would
introduce this capacitance. Resonant techniques (Section 4.2.2) are often necessary to
losses. The primary mechanisms of loss for a DC-DC converter are comprehensively
listed in Section 3.3.2 for PWM operation and in Section 3.4.2 for PFM operation. In
form factor, another challenge at lower voltage and current levels. As shown in Figure
1.1, for a fixed battery voltage, the value of filter inductance practically needed in a
8
Normalized Parameters
% Losses in L
4
1 2 3
Output voltage, Vo
Fig. 1.1: The effect of lower voltage and current on the external filter inductor.
1.3 Research Goals and Contributions 7
DC-DC converter design increases at lower voltages and currents, and the relative loss
due to the equivalent series resistance of the inductor also increases. High quality
inductors of large value and low current capability are an anomaly − their physical size
does not scale proportionally to their power handling. They are often not amenable to
planar configurations, and therefore, usually dominate the overall form factor of the
DC-DC converter. In Section 4.1.1 and Section 4.1.2, two circuit-level techniques are
shrink these inductors to chip-scale sizes, even they will require some measure of
The goal of this research is to design and implement DC-DC converters as low-
power and low-voltage enablers. This includes the development and demonstration of
DC-DC converters in nearly any portable electronic application. Several key research
• Developed a series of design techniques which decrease the size, cost, and energy
dissipation of low-voltage DC-DC converters. These include new ideas, such as:
reported output voltage and power levels: Greater than 70% efficiency at 0.2 V
build.
Several approaches to voltage scaling for low-power are reviewed. Aggressive voltage
Dynamic scaling of the voltage supply is proposed to trade performance for energy-
regulator topologies − buck, boost, and buck-boost − are introduced. The requirements
equations and closed-form expressions for losses are presented for both pulse-width and
prohibitive.
power system, individual control system, and circuit levels are presented which reduce
the overall size, cost, and energy dissipation of a single DC-DC converter, or an entire
5. Four key performance metrics are introduced, and their impact on dynamic DC-DC
converter design and the entire dynamic voltage scaling (DVS) system are discussed.
separate prototype converters. These prototypes were built to examine the feasibility of
the power system, control system, and circuit-level optimizations of Chapter 4, and to
directions.
10
Chapter 2
DC-DC Conversion as a
Low-Power Enabler
battery run-time. Perhaps the most effective low-power technique is to operate each
power dissipation with decreasing supply voltage. This comes at the expense of
run-time, dynamically adjusting the supply voltage to trade performance for energy
power design strategy assumes that the supply voltage is a free variable and can be set
to any arbitrarily low level with little penalty. In portable electronic systems, high-
techniques which are enabled by DC-DC converters, and the potential battery run-time
2.1 Voltage Scaling for Low-Power 11
2.1, the fundamental trade-off between the speed and power dissipation of a digital
objective are reviewed. Section 2.2 introduces the concept of dynamically scaling the
supply voltage to realize the speed-versus-power trade-off for systems with variable
proposed to dramatically decrease the power dissipation involved in driving the large
interface between the battery source and the load in a portable electronic system. A
that are enabled by simply regulating the battery source voltage with a DC-DC
converter.
2
E = C ⋅ V dd (Eq 2-1)
where C is effective capacitance fully charged and discharged over a voltage swing V dd ,
from a power supply of potential V dd . From (Eq 2-1), it is clear that a reduction of the
power supply voltage yields a quadratic savings in energy dissipation per computational
event.
illustrated in Figure 2.1, there is a fundamental trade-off between the energy consumed
With short channel MOS devices, carrier velocity saturation under high
voltages, there is little penalty in delay, but large potential power savings from supply
13
11
Normalized energy, delay
Vdd [Volts]
(around 0.7-0.9 V for the data in Figure 2.1) a large increase in circuit delay, with little
energy saving, is seen for a small decrease in supply voltage. It is in the region between
these two extremes that performance and energy consumption are readily traded if the
compromising computational throughput, operates the timing critical parts of the chip
at a high supply voltage, and reduces the voltage supply of the circuits not on the
critical path [Usami95], [Raje95], [Chang96], [Igarashi97]. This scheme, often called
clustered voltage scaling [Usami95], is conceptually illustrated in Figure 2.2. Here, the
speed critical circuitry is run at the high supply voltage, V ddH , while those circuits not
on the critical path are run at a lower supply voltage, V ddL . Communication from V ddL
For minimum power, greater than two separate voltages may be used per IC
[Chang96]. The primary limitation is the power introduced by the level converters.
VddH
Critical Path
VddH
Speed-critical
Circuitry
Arithmetic
Block
VddL VddH
Low-power Level
Circuitry Converter
Fig. 2.2: Conceptual illustration of using multiple supply voltages to reduce power dissipation.
2.1 Voltage Scaling for Low-Power 14
VddH
VddL
in 0
out
VddL VddH
in out 0
While each circuit block operated at lower voltage will effect some power savings, as
the number of separate voltage supplies increases, the overhead power of the additional
circuitry required to convert signals between these voltages begins to outweigh the
for increased silicon area by exploiting parallel and pipelined architectures. Hardware
may be duplicated to reduce the clock frequency of each processing element. This
allows the supply voltage to be scaled and often results in a significant reduction in
power dissipation.
input vectors, A, B, and C, are clocked into the datapath at a rate 1/T. The minimum
clock period, T, is set by the maximum propagation delay through the adder and
comparator. (The delay, set-up, and hold times of the registers are assumed negligible.)
The resulting output, (A+B) > C, is generated at the full throughput, 1/T. The total
2.1 Voltage Scaling for Low-Power 15
adder
A Σ
comparator
(A+B) > C
1/T
1/T
1/T
power dissipation is determined by the switching of the adder, comparator, and three
entire datapath is duplicated so that each may be clocked at a reduced frequency 1/2T.
This enables the supply voltage to be scaled, conserving power. However, the addition
of the multiplexer, clocked at the full throughput 1/T, does add some additional
overhead power. For the identical function, (A+B) > C generated at 1/T, the total power
dissipation is, in effect, now determined by the switching of an adder, comparator, three
registers, and a multiplexer at 1/T − switching over a supply voltage where the
maximum propagation delay through the adder and comparator is 2T. For example, if
the reference datapath is operated at 3.3 V, Figure 2.1 indicates that the parallel
implementation can run at 1.8 V. This yields a power dissipation of only 30% that of the
reference design.
samples are produced at a clock rate 1/T that is determined by the maximum delay
2.1 Voltage Scaling for Low-Power 16
adder
comparator
1/2T
A
1/2T
(A+B) > C
mux
B 1/T
1/2T
adder
comparator
C
1/2T
1/2T
1/2T
through either the adder or the comparator. This means that for a fixed throughput, 1/T,
the supply voltage can be scaled relative to the reference case, conserving power.
adder
A Σ
comparator
(A+B) > C
1/T 1/T
1/T
1/T
Since the energy per computational event ideally scales as V dd 2 while circuit
supply voltage of V dd = 1.5 V with V t = 1.0 V will have nearly identical performance to
the circuit running at V dd = 0.9 V will consume roughly one third the power.
combination of V dd = 0.9 V, V t = 0.5 V for a 20 MHz 16-bit ripple carry adder in a 1.2
µm CMOS process.
2.1.4 Discussion
However, the discussions above assume that the voltage supply is a free variable and
can be set to any arbitrarily low level with little penalty. In portable electronic systems,
sacrificed for lower power, rendering most voltage scaling techniques impractical for
dissipation with little sacrifice in peak performance. Since the processor spends a large
fraction of time idling, and performs mainly low throughput and high latency processes,
it can be shut down for the majority of its cycles, significantly reducing power. Two
Scaling, further decreases average power dissipation by reducing the energy per
2.2 Dynamic Voltage Scaling for Energy-Efficient GPP 19
operation of the lower throughput tasks − those tasks which otherwise dominate the
plotted versus time, and it is indicated that peak processor throughput (limited by the
peak performance of the processor) is demanded only a small percentage of the time.
The processor spends most of its time idling, and performs the majority of its cycles on
typical processor usage statistics to conserve power. To determine the relative merits of
(average energy per operation) x (minimum delay per operation) (Eq 2-2)
Compute-intensive and
low-latency processes
Desired Throughput
time
designs. Since the peak performance of the processor is, itself, a key specification,
average power and average energy per operation are poor metrics. Either can be reduced
at the expense of performance; the former by reducing the clock frequency; the latter by
reducing the voltage supply. For a fixed peak throughput, the minimum metric of (Eq 2-
2) indicates the largest number of operations that can be performed from a fixed battery
capacity − the most energy-efficient design. For a fixed number of operations, the
minimum metric indicates the maximum throughput of the processor − the highest-
performance design.
[Burd95]. A maximum clock frequency, f MAX = 100 MHz, is achieved at 3.3 V, where
the energy per operation is E MAX = 4.5 nJ. The relative delay and energy per operation
scale with voltage as shown in Figure 2.1. Although the processor must occasionally
deliver peak throughput to service certain operations, the majority of its energy is
consumed on low throughput and high latency processes. For the purposes of this
performed at 5 MHz, while the other 1% of the operations are performed at f MAX . The
The most obvious technique for reducing the power consumption of the
processor is to shut it down when it idles. In Figure 2.8, all operations are computed at
the maximum clock speed, f MAX . Lower throughput tasks are performed by waking the
Excess throughput
Peak
Throughput
Delivered
Desired
time
Fig. 2.8: Processor power management: Wake up → compute ASAP → sleep mode
[Burd96].
In the ideal case, the processor can shut down or wake up immediately and
with no energy overhead, and dissipates no power when it idles. In this way, the average
since the processor operates from a constant supply voltage, despite the fact that the
average power scales with decreasing throughput requirements, the energy per
operation is unchanged. Evaluation of the metric of (Eq 2-2) results in a figure of merit
equal to:
top of sleep mode. In this scheme, illustrated in Figure 2.9, the clock frequency of the
processor is reduced below f MAX to further decrease the average power dissipation.
Evaluation of the metric of (Eq 2-2), with a clock frequency reduction to f clk = f MAX / 2
results in:
Peak
Throughput fCLK Desired
Reduced Delivered
time
Comparison of (Eq 2-5) with (Eq 2-4) shows that this technique results in an even less
energy-efficient design.
reduce the average power dissipation of the processor, they do not take advantage of the
lower throughput requirements to scale the energy per operation. Because the majority
of operations are still performed on lower throughput tasks, the circuits usually
complete operations far faster than required, and according to the data in Figure 2.1, are
unnecessarily wasteful of energy. If instead, the clock and the voltage are dynamically
scaled together to meet the real-time computational demands of the user as in Figure
2.10, lower energy per operation can be achieved on the lower throughput tasks
This is shown in Figure 2.11, where the data in Figure 2.1 is redrawn to display
energy per operation versus delivered throughput. The gray line plots this data for a
fixed 3.3 V power supply voltage; for a fixed voltage supply, regardless of the
processor throughput, the energy per operation is unchanged. The solid black line
shows the same data for a scaled supply voltage − one that ensures that the circuit delay
2.2 Dynamic Voltage Scaling for Energy-Efficient GPP 23
Peak
Throughput
Delivered = Desired
time
Fig. 2.10: The clock and voltage are scaled dynamically.
~10x Energy
0.5 Reduction
Fig. 2.11: Energy per operation versus throughput for a digital CMOS circuit.
just meets the throughput requirements of the clock. At the 1.05 V operating point, a
9.9x improvement in energy per operation can be realized. While this requires a 20x
reduction in clock frequency, in many portable electronic systems, this operating point
yields sufficient throughput for the majority of operations. As a result, a nearly 9.9x
maximum throughput, f MAX = 100 MHz, is maintained at 3.3 V for the required 1% of
2.2 Dynamic Voltage Scaling for Energy-Efficient GPP 24
the operations. At this operating point, the energy per operation is 4.5 nJ. However, the
remaining 99% of the operations require computation at only 5 MHz, allowing a circuit
delay of twenty times the 10 ns minimum. From Figure 2.1, the processor can achieve
this throughput from a 1.05 V supply, yielding a reduction in energy per operation to
and
EAVE = (0.99) (0.4 nJ) + (0.01) (4.5 nJ) = 0.44 nJ (Eq 2-7)
management techniques.
2.2.3 Discussion
conventional static DC-DC converter. Whereas a static DC-DC converter must maintain
energy is consumed on low throughput and high latency processes. Otherwise, DVS
effects no substantial energy savings. In addition, the energy saved by DVS must be
conserved by the dynamic DC-DC converter. This means that adaptations in the output
2.3 Low-Swing Interconnect 25
voltage must be energy efficient, and since the majority of energy in a DVS system is
Chapter 5 details these and other DVS system and circuit-level considerations.
The power dissipation associated with driving large capacitive off-chip busses
Dynamic voltage scaling has been proposed to reduce the energy consumption of the
major components of this subsystem − the processor core and the memory ICs. The
resulting system is expected to consume no more than 450 mW at 100 MIPS and 3.3 V,
MIPS and 1.05 V). However, these figures neglect the dissipation associated with
interchip communication.
per bit, at the full system throughput. Assuming an activity factor of 25%, if each bit is
fully driven from rail-to-rail, the associated power dissipation would be:
2
P bus = ( 32 bits ) ⋅ ( 50 pF ) ⋅ ( 0.25 ) ⋅ ( 3.3 V ) ⋅ ( 100 MHz ) = 435 mW (Eq 2-9)
2
P bus = ( 32 bits ) ⋅ ( 50 pF ) ⋅ ( 0.25 ) ⋅ ( 1.05 V ) ⋅ ( 5 MHz ) = 2.2 mW (Eq 2-10)
2.3 Low-Swing Interconnect 26
in the most energy-efficient mode. In both cases, this approximately doubles the power
[Yamauchi94], have been proposed for this problem. While many of these techniques
have been demonstrated with some success, they either add too much complexity to the
system, or are not as conservative with power as they might be. An alternative scheme
is proposed here.
Voltage scaling for low-power is the underlying concept of the low-swing I/O
bus transceivers [Burd95]. Figure 2.12 shows a block diagram of the approach. The
voltage, V LO . The gates of NMOS buffer devices M1 and M2 are driven at full-rail
VDD
VDD VLO
VDD VLO / 2
VLO -
0 VDD
+
0
50pF 0
IC IC
VLO
VDD in
M1
VDD out
0
VLO
M2 0
Since they drive their large output load capacitance between only 0 V and V LO , power
between the low-voltage rails. In the ideal case, the power dissipation of the receivers
2
reduced by the factor ( V dd ⁄ V LO ) . In [Burd95], a 200 mV signal swing has been
proposed, and a test chip verified successful operation above 100 MHz [Burd98]. The
new bus transceiver system reduces this component of power dissipation to:
a factor of 272 and 27.5 lower than the figures reported in (Eq 2-9) and (Eq 2-10),
respectively − and nearly negligible compared to the power dissipation of the processor.
This low-swing bus architecture has two distinct advantages over existing
supply to the drivers so that, unlike other low-swing I/O architectures that employ
linear regulators, the majority of the power saved by the transceiver circuitry is not
dissipated in the regulator. Second, this approach uses single-ended, rather than
differential, signals. This means that pin count and board-level routing complexity are
differential architectures.
2.3.1 Discussion
challenging: 80 µW at 0.2 V is far lower in voltage and power than any previously
reported converter. However, since the power savings are so large as to make power
dissipation nearly negligible, the efficiency need not be as aggressively high as in most
since V LO need not be tightly regulated − it must be some voltage which is much
smaller than V DD − some compromises can be made in the design of the converter.
Voltage regulation as an interface between the battery source and load can
further enhance system run-time. A circuit may be designed such that its optimum
operating voltage is the end-of-life voltage of a specific cell, apparently minimizing its
power consumption without the use of a DC-DC converter. This not only makes the
circuit design challenging (the voltage of a typical AA-type lithium ion cell may vary
by as much as ± 20% of its nominal value throughout its discharge), but because the cell
discharge characteristic is not flat, the circuit will consume greater than its minimum
operating power from the cell throughout the majority of its discharge. If a DC-DC
converter is inserted between the cell and the load, and the converter’s output voltage is
maintained down to the end-of-life cell voltage, the circuit will consume its minimum
operating power independent of the cell voltage, substantially extending system run-
time (by as much as 50% for a digital CMOS circuit powered by a single lithium ion
cell).
Figure 2.13 shows typical low-rate battery discharge curves for three
4.5
Li Ion
NiMH
1.5
NiCd
0
0 300 600 900
Charge Delivered q [mAh]
Fig. 2.13: Typical low-rate discharge characteristics for AA-type Nickel Cadmium (NiCd),
Nickel Metal Hydride (NiMH), and Lithium Ion (Li Ion) cells. Data is approximated from
[Caruthers94].
Nickel Metal Hydride (NiMH), and Lithium Ion (Li Ion). Consider a block of
throughput-constrained logic run directly from a NiMH cell and designed to operate
down to the end-of-life cell voltage. If the power consumption of the logic is dominated
by the dynamic component, and the circuitry is clocked at a frequency f 0.9 to meet
throughput constraints at the minimum cell voltage v ( q ) = 0.9 V , then the circuitry will
2
P L(min) = f 0.9 ⋅ C eff ⋅ 0.9 (Eq 2-13)
Here, C eff is the effective switching capacitance (commonly expressed as the product of
points q in the cell discharge characteristic v(q), the power consumption of the circuitry
is given by:
2
2 v(q)
P L ( q ) = f 0.9 ⋅ C eff ⋅ v ( q ) = P L ( min ) ⋅ ------------- (Eq 2-14)
2
0.9
2.4 Voltage Regulation Enhances Battery Run-Time 30
At initial cell voltage, this is a factor of 2.78 times P L(min), and at nominal cell voltage,
a factor of 1.78 times P L(min) . Thus, the load is seen to consume greater than minimum
P out
η ≡ ---------- (Eq 2-15)
P in
and zero dropout voltage is inserted between the battery and the load, and the output of
the converter is regulated to the end-of-life cell voltage, the logic consumes P L(min)
independent of the cell voltage, and the power drawn from the cell at any point q in its
P L ( min )
P ( q ) = ------------------- (Eq 2-16)
η
DC-DC conversion on system run-time. This analysis considers analog circuitry with
compares system run-time when these loads are run directly from the battery source,
and from the battery source at a minimum voltage through a linear regulator or a
switching regulator.
its cell voltage v(q) after a charge, q, has been delivered to the load. At full capacity
( q = 0 ), the cell has an initial voltage v ( 0 ) = V 1 . The nominal cell voltage lies in the
V1 Initial
Nominal
V2
V3
0 Q1 Q2 QA
Fig. 2.14: A piecewise linear model of a typical low-rate cell discharge characteristic.
full capacity, E A , is the area under the entire discharge curve. The mean cell voltage
found by solving the following differential equation which governs the cell discharge at
·
q = i(q) (Eq 2-17)
q = 0, t = 0 (Eq 2-18)
yielding:
2.4 Voltage Regulation Enhances Battery Run-Time 32
QA
∫ ---------
dq
tA = (Eq 2-19)
i(q)
0
Figure 2.15 shows the three loads considered in this analysis, (a) a constant
current load I, (b) a resistive load R, and (c) a constant power load P, each attached
across the terminals of a cell whose discharge characteristic v(q) is described by Figure
2.14.
In Figure 2.15a, the current drawn from the battery is constant and equal to I.
QA
t A = -------- (Eq 2-20)
I
v(q)
i ( q ) = ----------- (Eq 2-21)
R
+ + + Constant
v(q) v(q) R v(q) Power
I Load, P
- - -
Fig. 2.15: Battery loading conditions: (a) a constant current load I, (b) a resistive load R, (c)
a constant power load P.
2.4 Voltage Regulation Enhances Battery Run-Time 33
proves ungainly and provides little insight. However, if the simplifying assumption that
the mean load current, averaged over the system run-time ( t ∈ [ 0, t A ] ) is equal to the
v(q )
i ( t ) = i ( q ) = ----------- (Eq 2-22)
R
QA ⋅ R
t A = ---------------- (Eq 2-23)
v( q)
Since the cell voltage v(q) is relatively flat during the majority of the cell discharge, the
approximation of (Eq 2-22) is valid for any of the discharge characteristics of Figure
In Figure 2.15c, the load draws a constant power P from the cell, such that:
P
i ( q ) = ----------- (Eq 2-24)
v(q)
and:
EA QA ⋅ v( q )
t A = ------- = ----------------------- (Eq 2-25)
P P
QA
t Ao = -------- (Eq 2-26)
I
In the idealized case, the linear regulator has a dropout voltage of zero and a
quiescent operating current which is negligible with respect to I (see Section 3.6.1).
Thus, the supply may be regulated to the minimum voltage, V min ≤ v ( q ) , at which the
load can operate, minimizing its power consumption, and the quiescent current of the
regulator may be ignored. However, because the same current I drawn by the load flows
through the regulator, the power which is conserved by running the load at V min is
dissipated in the regulator. (The dissipation in the regulator is I ⋅ V min .) The battery
tA
-------
- = 1 (Eq 2-27)
t Ao
efficiency η, the load consumes a constant and minimum power. The power drawn from
P L ( min ) I ⋅ V min
P = ------------------- = ------------------- (Eq 2-28)
η η
Substituting (Eq 2-28) into (Eq 2-25), and normalizing with respect to t Ao gives:
2.4 Voltage Regulation Enhances Battery Run-Time 35
tA η ⋅ EA η ⋅ v ( q -)
-------
- = ------------------------
- = ------------------ (Eq 2-29)
t Ao V min ⋅ Q A V min
throughput constraints at the minimum voltage V min , and that has an effective switching
1
R eff = --------------------------------- (Eq 2-30)
f V ( min ) ⋅ C eff
Substitution of (Eq 2-30) into (Eq 2-23) gives the baseline system run-time:
Q A ⋅ R eff
t Ao = ---------------------- (Eq 2-31)
v(q)
If the load is run from the minimum voltage V min at which throughput
V min
I min = ------------ (Eq 2-32)
R eff
which is sourced through the regulator from the battery source. This current represents
the minimum operating current of the load. Substitution of I min in (Eq 2-32) for I in (Eq
tA
-------
EA
- = ------------------------ v ( q )-
- = ----------- (Eq 2-33)
t Ao V min ⋅ Q A V min
2
V min
P = P L ( min ) = ------------ (Eq 2-34)
R eff
which represents the minimum operating power of the load. The average power drawn
2
P L ( min ) V min
P = ------------------- = ------------------ (Eq 2-35)
η η ⋅ R eff
and:
2
tA η ⋅ EA η ⋅ v(q)
2
-------- = -------------------------------- = ---------------------- (Eq 2-36)
t Ao 2 2
( V min ⋅ Q A ) V min
2.4.5 Results
is the ratio of the mean cell voltage (averaged over the delivered charge, q) to the
minimum voltage required by the load. For convenience in summarizing the results, the
v(q )
β ≡ ------------ (Eq 2-37)
V min
2.4 Voltage Regulation Enhances Battery Run-Time 37
In terms of β, Table 2.1 gives the run-time enhancement factor, K, for a linear
the run-time relative to the baseline run-time when the load is run directly from the
battery source,
tA
K ≡ -------- (Eq 2-38)
t Ao
Figure 2.16 shows the system run-time enhancement for NiCd, NiMH, and Li
Ion cells loaded with analog and digital circuitry achieved by simply regulating the
battery source voltage with a linear regulator, and a 90% and 100% efficient DC-DC
converter. Here, the output voltage of each converter is maintained at the end-of-life
cell voltage.
The results shown in Table 2.1 can be used to predict the benefits of different
system run-time for a constant-current load (e.g. many analog circuits). It should only
be used if a stabilized voltage improves the performance of the load circuitry. With a
digital CMOS load, the linear regulator provides an improvement by the factor β.
Regardless of the load type, a switching regulator results in a value of K which is that
for a linear regulator, multiplied by an additional factor ηβ. As long as the efficiency of
the regulator is high enough that ηβ > 1, the switching regulator will give a longer run-
The benefits of a switching regulator are greatest where β is large; that is,
where the minimum required load voltage is small compared to the average battery
voltage. This makes intuitive sense, since an unnecessarily high voltage is wasteful of
energy. With a load that is designed to run down to the end-of-life cell voltage, the
factor β is only a function of the battery characteristic, and, for the discharge
characteristics of Figure 2.13, is 1.33 for NiMH or NiCd cells, and 1.26 for Li Ion.
Note, however, that for a load with a minimum operating voltage below the end-of-life
voltage of its battery source, β can be much higher. For example, consider the low-
operate at a 1.1 V minimum supply voltage, were run from a Li Ion cell, β would be
3.27. In this system, even a very low efficiency switching regulator would be desirable
− even with 31% efficiency, it would out-perform an ideal linear regulator. Efficiency is
still important, however − in all cases, the run-time with a DC-DC converter is directly
2.0
β = 1.33 β = 1.33 β = 1.26 Analog Load
System Run-time Enhancement, K
Linear Reg
Analog Load
1.5 DC-DC (η=0.9)
Analog Load
DC-DC (η=1.0)
1.0
Digital Load
Linear Reg
Digital Load
0.5
DC-DC (η=0.9)
Digital Load
DC-DC (η=1.0)
0.0
Fig. 2.16: Battery run-time enhancement achieved by regulating the battery source voltage to
the end-of-life cell voltage.
2.4 Voltage Regulation Enhances Battery Run-Time 39
proportional to the efficiency of the converter. In this example, with 90% efficiency, as
is readily achieved using the design techniques presented in Section 4.2, the system
run-time would be 9.64 times longer than if the chipset were run directly from the Li
same enhancement of run-time may also be achieved by simply increasing the capacity
of the battery source. The battery is often the physically largest and most expensive
volume, and complexity of the design. Thus, from a system design standpoint, it is
important to compare the volume required for the converter to the volume that would be
converter. The volume of the converter needed to achieve this enhancement, ∆S DC-DC ,
may be estimated from the power it supplies, P L(min) , and its power density, D P(DC-DC) :
P L ( min )
∆S DC-DC = -------------------------- (Eq 2-39)
D P ( DC-DC )
the battery capacity would need to be increased by the factor K. The resulting increase
∆S B = S B0 ( K – 1 ) (Eq 2-40)
where ∆S B is the volume of the additional battery capacity, and S B0 is the initial battery
volume. The initial battery volume may be calculated from the energy it stores at full
EA
S B0 = ------------------ (Eq 2-41)
D E ( bat )
The volume of the DC-DC converter is related to the load power, as illustrated
by (Eq 2-39), whereas the volume of the additional battery capacity is related to the
integral of the load power − the total energy consumed by the load over the system run-
time. These two quantities can only be compared by specifying the enhanced run-time,
t A . In the case that a DC-DC converter is used, the load on the battery is a constant
P L ( min ) ⋅ t A
E A = ----------------------------- (Eq 2-42)
η
Substituting this expression into (Eq 2-41), and the result into (Eq 2-40), gives
t A ⋅ P L ( min ) ( K – 1 )
∆S B = ----------------------------- ⋅ ------------------ (Eq 2-43)
D E ( bat ) η
∆S B D P ( DC-DC ) ⋅ t A ( K – 1 )
---------------------- = ------------------------------------ ⋅ ------------------ (Eq 2-44)
∆S DC-DC D E ( bat ) η
Conceptually, (Eq 2-44) compares the energy density of the battery (D E(bat) ) to
the effective energy density of the converter − the factor D P ( DC-DC ) ⋅ t A gives the
energy handled by the converter per volume, and the factor ( K – 1 ) ⁄ η corrects this for
the amount of energy savings the converter effects, relative to the amount of energy it
handles. Although the position of η in (Eq 2-44) is at first counter-intuitive, recall that
∆S B D P ( DC-DC )
---------------------- = -------------------------- ⋅ t A ⋅ ( K 0 – 1 ⁄ η ) (Eq 2-45)
∆S DC-DC D E ( bat )
Since K 0 is equal to β or β 2 (see Table 2.1), the ratio, (Eq 2-45), is seen to increase with
converters with power densities above 1 W/cm 3 can be achieved through the use of the
techniques discussed in Section 4.1. Using these power and energy densities in
conjunction with (Eq 2-44), it is possible to evaluate the relative converter or additional
For example, again consider the system introduced in Section 2.4.5. There, it
was shown that a 90% efficient DC-DC converter with a regulated 1.1 V output can be
used to enhance system run-time from a Li Ion source by a factor of K = 9.64 . For an
8 h target run-time, the volume required by 8.64 times more Li Ion capacity is roughly
256 times greater than that required by the converter. If a shorter run-time is targeted,
the additional battery volume needed to achieve the same percentage of enhancement is
smaller, but, because its power handling requirements are unchanged, the volume of the
DC-DC converter remains the same. Thus, for short run-times, adding battery capacity
requires less volume than adding a DC-DC converter. However, based on the same
factors of this example, for any run-time longer than two minutes, the additional battery
It may be concluded that, with the exception of systems designed for very short
involve only a small increase in volume, much smaller than the increase in battery
Chapter 3
DC-DC Converter
Fundamentals
expressions for losses are presented for the three basic low-voltage CMOS switching
double-pole switch chops Vin producing a rectangular wave having an average voltage
equal to the desired output voltage. A low-pass filter passes this DC voltage to the
output while attenuating the AC ripple to an acceptable value. The output is regulated
by comparing Vo to a reference voltage, Vref , and adjusting the fraction of the cycle for
which the switch is shorted to Vin . This pulse-width modulation (PWM) controls the
3.1 Introduction to Switching Regulators 43
average value of the chopped waveform, and thus controls the output voltage. Unlike a
which approaches 100% as the components are made more ideal. In practice,
efficiencies above 75% are typical, and efficiencies above 90% are attainable.
There are several simple alternative arrangements of the switching and filter
components that can be used to produce an output voltage larger or smaller than the
input voltage, with the same or opposite polarity. Some of these will be discussed
below. However, many of the design issues are similar, so first one topology, the step-
The power train of the low-output-voltage buck circuit, which can produce any
arbitrary output voltage 0 ≤ V o ≤ V in , is given in Figure 3.2. The basic PWM operation
is as follows: The power transistors (pass device M p and rectifier M n ) chop the battery
input voltage Vin to reduce the average voltage. This produces a square wave of variable
duty cycle D and constant period Ts = f s -1 at the inverter output node, v x . A typical
periodic steady-state v x (t) waveform is shown in Figure 3.3. The second-order low-pass
3.1 Introduction to Switching Regulators 44
+
Mp iLf
Vin Cin
+ Lf +
Mn vx Cf Vo
- - -
PMOS
Vin
on
NMOS vx(DC) = Vo
on
0
DTs (1-D)Ts
filter (L f and C f ) passes the desired DC component of this chopped signal, while
attenuating the AC to an acceptable ripple value. In the ideal case, the DC output
voltage is given by the product of the input voltage and the duty cycle:
V o = V in ⋅ D (Eq 3-1)
duty cycle of the rectangular wave at v x , and ultimately, the DC output voltage, to
compensate for input and load variations. The pulse-width modulation is controlled by
a negative feedback loop, shown in the block diagram of Figure 3.1, but omitted from
Figure 3.2 for simplicity. Some detail on ultra-low-power PWM design is included in
Chapter 4.
3.2 DC-DC Requirements in Portable Systems 45
multimedia Infopad terminal, the six voltage converters are the dominant source of
hardware: Unused circuitry is powered-down and gated clocks are employed to reduce
power consumption during idle mode [Chandrakasan94b]. Such techniques may present
severe load variations (up to several orders of magnitude), and the system may idle for
a large fraction of the overall run-time. This implies the need for a high conversion
efficiency not only under full load, but over a large load variation. Furthermore, in the
power (control power) of the regulator must be kept to an even lower level to ensure
that it does not contribute significantly to the overall dissipation. For example, a
I/O, pen input and full motion video, and consumes less than 5 mW at 1.1 V. The
control circuit for a converter supplying this chipset must have substantially lower
quiescent power.
techniques at the power system and circuit levels to improve the energy efficiency of
these converters. At the power system level, resource sharing between converters is
used to minimize control system overhead. Low-voltage digital control which exploits
power train circuit optimizations for high efficiency at ultra-low output voltages are
presented.
greater variety of technologies are integrated into a single system, their voltage
conversion needs grow. While successive generations of high performance digital ICs
continue to require higher voltages for headroom and signal distortion considerations.
performance DC-DC converter, including the IC and all external components, can cost
3.2 DC-DC Requirements in Portable Systems 47
as much as nine dollars 1 . Since as many as six DC-DC converter outputs may be
required in a portable electronic device [Truman98], the overall power system may
reduce the cost of the power system. Current-day DC-DC converters require as many as
ten external components. The design methodology presented in Chapter 4 reduces this
number to three: One input bypass capacitor, and an output filter inductor and capacitor.
In addition, the methodology allows for the integration of several power supplies on a
single IC, further reducing cost. Finally, since vanilla digital CMOS integration is
proposed, small custom power supplies can be integrated together with their own digital
CMOS loads.
each DC-DC converter output requires its own external filter elements. These
operating frequencies are proposed to reduce the values of these elements, thereby
mass. Since several DC-DC converters are required in almost any portable electronic
device, minimization of the physical size of each is a key design objective. The six
voltage converters in the Infopad terminal consume 12% of the printed circuit board
surface area [Truman98]. In addition, the large inductors in DC-DC converters often
1. Cost of the MAX887 purchased through a distributor in volumes of 1000, including 10 external compo-
nents.
3.2 DC-DC Requirements in Portable Systems 48
The techniques described in Section 3.2.2 for power system cost reduction are
equally effective in reducing overall power system size. Higher levels of functional
power supplies on a single IC, and power supplies together with their loads reduces the
Chapter 4, converter topology and battery voltage choices can have a profound impact
on the size of the overall power system. In addition, since there is a fundamental trade-
off between the size of a DC-DC converter and its losses (see Section 4.1.1) the size
and efficiency of different converters in the system may be traded to yield the optimum
physical size of a DC-DC converter. Higher operating frequencies reduce the required
values of inductance and capacitance, and ideally, their form factor. “Minimum
inductor” designs yield the minimum form factor inductor for a given application.
manufacturers use linear, rather than switching regulators for all DC-DC down-
The converters are used only in PWM mode, where the switching frequency (and
therefore, the frequencies of fundamental and harmonic switching noise) is known. The
switching frequency is chosen so that the higher-order harmonics are kept outside of the
3.3 PWM Operation 49
performance. The magnitude of the noise is reduced with careful physical design. All
power traces in the PCB are kept short and wide, minimizing the area, and thus the stray
inductance, in all critical high current loops. A closed-core output filter inductor design
offers a closed magnetic path to contain flux. Finally, a more recent innovation called
emissions.
Figure 3.5 shows the steady-state operating waveforms of the buck circuit in
PWM operation. The switching cycle is initiated when PMOS device, M p , turns on.
During the interval, D, of the switching period, Ts , the inverter output node, v x , is
shorted to Vin . A constant positive potential, Vin -Vo , is applied across the inductor, and
i Lf linearly increases from its minimum value to its maximum value. Some of the
energy removed from the battery is stored in the magnetic field of the inductor, and
Then, the PMOS device is turned off, and the NMOS rectifier device, M n , is
turned on to pick up the inductor current, shorting v x to ground. During this interval,
(1-D) of the cycle, a constant negative potential is applied across the inductor, and i Lf
linearly decreases from its maximum value to its minimum value. Excess energy in the
inductor is delivered to the output filter capacitor and load. The cycle then repeats by
the battery during a switching period is equal to the charge consumed by the load.
3.3 PWM Operation 50
Vin Ts
vx (t) 0
Io
iLf (t) ∆I
0
Vin
vgp (t) 0
Vin
vgn (t) 0
d Ts (1-d) Ts
Fig. 3.5: Periodic steady-state PWM waveforms for the buck circuit.
In Figure 3.6, the rectangular wave of the inverter output node is applied to the
second order low-pass output filter of the buck circuit (L f and C f ) which passes the
ripple value. Load R L draws a DC current I o from the output of the filter. Figure 3.7
shows the nominal steady-state i Lf(t) and v o (t) waveforms for a rectangular input v x (t).
iLf
Io
vx (t) +
Lf +
Cf Vo RL
(D, fs)
-
-
Fig. 3.6: The output filter of the buck circuit (Lf and Cf) with load RL.
3.3 PWM Operation 51
Ts
iLf (t)
∆I Io
vo (t)
∆V Vo
Fig. 3.7: Nominal steady-state waveforms of the buck circuit output filter.
–2
L f ⋅ C f » ω s , where ω s = 2πf s , and f s is the switching frequency of the converter. In
this case, the filter components may be sized independently, using time domain
analysis, rather than frequency domain analysis. Neglecting the effects of output
symmetric about the average load current I o . The peak-to-peak current ripple may be
found by integrating the AC component of the v x (t) waveform over a fraction, D, of one
cycle, yielding:
V in ⋅ D ⋅ ( 1 – D ) Vo ⋅ ( 1 – D )
∆I = ---------------------------------------- = ----------------------------- (Eq 3-2)
Lf ⋅ fs Lf ⋅ fs
The output filter capacitor is selected to ensure that its impedance at the
switching frequency, including its equivalent series resistance (ESR), is small relative
3.3 PWM Operation 52
to the load impedance. Thus, the AC component of the inductor current flows into the
filter capacitor, rather than the load. For many capacitor technologies at frequencies
above several hundred kilohertz, the resistive impedance dominates over the capacitive
reduce both output voltage ripple and conduction loss (see below). For this reason, a
high-Q capacitor technology, such as multilayer ceramic, is typically used, and even at
Considering only capacitive impedance, the peak-to-peak output voltage ripple may be
found through charge conservation. Assuming the AC inductor current flows only into
∆I Vo ⋅ ( 1 – D )
∆V = ---------------------- = -------------------------------- (Eq 3-3)
8 ⋅ Cf ⋅ fs 8 ⋅ Lf ⋅ C f ⋅ f s
2
This output voltage ripple is symmetric about the desired DC output voltage
Vo , and, for the v x (t) waveform shown in Figure 3.7, is piecewise quadratic with period
Ts .
(Eq 3-2) and (Eq 3-3) illustrate the two principle means of miniaturizing a DC-
DC converter. First, it can be readily seen that the necessary values of filter inductance
–1
and capacitance decrease with f s . Thus, a higher operating frequency typically results
ripple, it is the L f ⋅ C f product, rather than the values of the individual components, that
is important. Through choice of a higher current ripple, ∆I, a lower filter inductance
The power train of the low-output-voltage buck circuit, including all series
power transistors, is shown in Figure 3.8. Listed below are the chief sources of
dissipation that cause the conversion efficiency of this circuit to be less than unity. In
2
P q = i rms ⋅ R (Eq 3-4)
where i rms is the root mean squared current through the component, and R is the
2 2 2
i rms = i rms ( DC ) + i rms ( AC ) (Eq 3-5)
where:
Ls
+
Rbat RCin Mp iLf
Vgp Io
Rs
Vin Cin
Lf RCf +
Vgn
Mn Cx Cf Vo RL
- -
2 2
i rms ( DC ) = d ⋅ I o (Eq 3-6)
and
1 ∆I 2
i rms ( AC ) = d ⋅ --- ⋅ ------
2
(Eq 3-7)
3 2
Here, 0 ≤ d ≤ 1 is a weighting factor which indicates the duty cycle of current flow
through the component, I o is the DC load current, and ∆I is the peak-to-peak inductor
current ripple.
AC conduction loss is a fixed quantity and may substantially degrade efficiency at light
load.
Raising and lowering the gate of a power transistor each cycle dissipates an
average power:
Pg = Eg ⋅ fs (Eq 3-8)
where E g is directly proportional to the gate energy transferred per off-to-on-to-off gate
transition cycle (which can include some energy due to Miller effect), and includes
Gate-drive loss is independent of load current and will therefore degrade light-
load efficiency.
switching of the power MOSFETs are described below. Each is independent of load.
3.3 PWM Operation 55
A short-circuit path may exist temporarily between the input rails during
necessary to provide dead-times in the conduction of the MOSFETs to ensure that the
If the durations of the dead-times are too long, the body diode of the NMOS
power transistor may be forced to pick up the inductor current for a fraction of each
cycle. Since in low-voltage applications, the forward bias diode voltage ( V d ≈ 0.7 V )
can be comparable to the output voltage, its conduction loss may be significant:
where t err is the timing error between complementary power MOSFET conduction
intervals.
Furthermore, when the PMOS device is turned on, it must remove the excess
minority carrier charge from the body diode, dissipating an energy bounded by:
E rr = Q rr ⋅ V in (Eq 3-10)
1 2
P Cx ( LH ) = --- ⋅ C x ⋅ V in ⋅ f s (Eq 3-11)
2
3.3 PWM Operation 56
some or all of the gate-drain overlap (Miller) capacitance C gd of the power transistors,
wiring capacitance from their interconnection, and stray capacitance associated with L f .
associated with the connection of an off-chip filter inductor, which include a bond pad,
When M p is turned off, the inductor begins to discharge C x from Vin to ground.
NMOS device is turned on too late, v x will be discharged below ground, until the body
diode is forced to conduct (see above). If the NMOS device is turned on too early, it
1 2 1 2
P Cx ( HL ) = --- ⋅ C x ⋅ v x ⋅ f s ≤ --- ⋅ C x ⋅ V in ⋅ f s (Eq 3-12)
2 2
Energy storage by the stray inductance L s in the loop formed by the input
decoupling capacitor C in and the power transistors causes dissipation (Figure 3.9).
iLs φ
Imax
Ls i(t)
Mp Imin
φ i (t) iLs(t)
Mn
φ ELs = 0 ELs = 1/2 Ls Imax2
i (t)
ELs = 1/2 Ls Imin2 ELs = 0
Here, M p and M n are modeled as ideal switches, and L f is modeled as a current source
1 2 2
P Ls = --- ⋅ L s ⋅ ( I min + I max ) (Eq 3-13)
2
∆I
I min = Io – ------ (Eq 3-14)
2
and
∆I
I max = I o + ------ (Eq 3-15)
2
layout, and is reduced by minimizing the area of this critical high current loop. In a
a conductor that overlaps a return path in a different layer, with thin dielectric
The PWM and other control circuitry consume static power. In low-power
applications, this control power may contribute substantially to the total losses, even at
full-load.
3.4 PFM Operation 58
While a PWM DC-DC converter can be made to be highly efficient at full load,
many of its losses are independent of load current, and it may, therefore, dissipate a
significant amount of power relative to the output power at light loads. Figure 3.10
plots total losses versus a 1000:1 load range for a typical PWM buck converter. As the
load scales downward, AC conduction loss, switching loss, and PWM control power
fixed minimum power dissipation. From this plot, it may be concluded that a PWM
converter which is 94% efficient at full load is roughly 3% efficient at one thousandth
full load. If the converter is used at full load for little of its operating time, energy loss
at light load will be the dominant limitation on battery run-time, and improving
5.5
5
Pdiss [% of full load]
4.5
3.5
3
−3 −2 −1 0
10 10 10 10
Pout [normalized to full load]
ACTIVE
V+
vo (t)
VREF
V-
One control scheme which achieves high efficiency over a wide load range is
3.11, the converter is operated only in short bursts at light load. Between bursts, both
power FETs are turned off, and the circuit idles with zero inductor current. During this
period, the output filter capacitor sources the load current. When the output is
discharged to a certain threshold below V REF, the converter is activated for another
burst, returning charge to C f . Thus, the load-independent losses in the circuit are
reduced. As the load current decreases, the idle time increases. Regulation is
maintained when the charge delivered through the inductor is equal to the charge
One major drawback of PFM control is that the switching period (the time
between charge bursts) is a function of load. Thus, the converter appears almost chaotic
periods of radio inactivity by tying the converter’s operating mode to the pulse timing
of the radio. For example, during the page/scan mode of a TDMA RF system, the
converter can be commanded into PWM mode, where the spectrum of the switching
3.4 PFM Operation 60
noise is well-controlled, while the receiver is active. During the periods of receiver
inactivity, the converter can be commanded into PFM mode for high energy-efficiency.
Figure 3.12 shows the steady-state buck circuit waveforms under PFM control.
The PFM operation is described heuristically in Figure 3.11: When the output voltage
waveforms of Figure 3.11, here, only a single switching event of the DC-DC converter
is used to deliver each burst of charge.) This charge burst is delivered with high energy
efficiency through the inductor as follows: The PMOS device is turned on for a time
interval, T pmos . Some of the energy removed from the battery is delivered to the output;
the rest is stored in the inductor. During this interval, the inductor current slews at a
rate of:
di L ( V in – V o )
-------- = -------------------------
- (Eq 3-17)
dt Lf
and reaches its peak value of i Lf = I p at the conclusion of the PMOS conduction
interval. The PMOS device is then turned off, and after a short dead-time, the NMOS
vo (t)
VREF
T Vin
vx (t)
Vo
0
Tidle
Tnmos Tpmos
Ip
shorted to ground, and the energy stored in the inductor is released to the output. The
di L –V
-------- = ---------o- (Eq 3-18)
dt Lf
The NMOS device is (ideally) turned off when i Lf decays to zero. At this time,
v x will ring up to Vo , and the circuit will idle with zero inductor current and the output
The total charge delivered through the inductor by each PFM burst is found by
integrating the area under the i Lf (t) waveform for one switching cycle of the DC-DC
converter:
1
Q L = --- ⋅ I p ⋅ ( T pmos + T nmos ) (Eq 3-19)
2
controlling variable is the PMOS conduction interval, T pmos . The NMOS conduction
equating the products of the linear inductor current slopes and the conduction intervals
( V in – V o ) ⋅ T pmos V o ⋅ T nmos
I p = ---------------------------------------------- = -------------------------- (Eq 3-20)
Lf Lf
( V in – V o )
T nmos = -------------------------- ⋅ T pmos (Eq 3-21)
Vo
2
1 T pmos ⋅ ( V in – V o ) ⋅ V in
Q L = --- ⋅ ----------------------------------------------------------- (Eq 3-22)
2 Vo ⋅ Lf
QL = Io ⋅ T (Eq 3-23)
where
Inductor Value
T pmos ⋅ ( V in – V o )
L f = ---------------------------------------------- (Eq 3-25)
2 ⋅ Io ( max )
As indicated by (Eq 3-22), a smaller value of inductance than that given in (Eq 3-25)
will support a larger load current, and will support I o(max) with a larger time between
pulses, Tidle .
Capacitor Value
The capacitor is selected to ensure that the peak-to-peak output voltage ripple,
calculated assuming that all of the charge delivered through the inductor is absorbed by
Cf:
3.4 PFM Operation 63
QL
∆V = ------- (Eq 3-26)
Cf
Section 3.3.2 for PWM operation. However, PFM converters are shut down during the
idle time, Tidle , between pulses and, with the exception of some static dissipation in the
control circuits, dissipate energy only during pulses. Thus, the analysis below presents
The overall efficiency of the converter in PFM operation is then expressed as the ratio
given by:
E pulse
η = ---------------------------------- (Eq 3-28)
E pulse + E diss
T pulse
∫
2
Eq = i ( t ) R dt (Eq 3-29)
0
3.4 PFM Operation 64
where i(t) is the current through the component, T pulse = T pmos + T nmos , and R is the
Raising and lowering the gate of a power transistor each cycle dissipates an
energy E g . This is directly proportional to the gate energy transferred per off-to-on-to-
off gate transition cycle (which can include some energy due to Miller effect), and
PMOS Turn-On
The power PMOS device is always turned on with the converter idling − in
steady-state, v x = Vo and i Lf = 0. The energy stored on C x just prior to PMOS turn-on is:
1 2
E Cx ( initial ) = --- ⋅ C x ⋅ V o (Eq 3-30)
2
The PFM switching cycle is initiated when M p charges C x from v x = Vo to v x = Vin . The
1 2
E Cx ( final ) = --- ⋅ C x ⋅ V in (Eq 3-31)
2
The energy drawn from the battery during this transition is equal to:
E in = V in ⋅ ∆Q Cx = V in ⋅ C x ⋅ ( V in – V o ) (Eq 3-32)
1 2
E Cx ( IH ) = E in – ( E Cx ( final ) – E Cx ( initial ) ) = --- ⋅ C x ⋅ ( V in – V o ) (Eq 3-33)
2
3.4 PFM Operation 65
The PMOS off to NMOS on transition is nearly identical to that in PWM mode
too short, M n discharges C x through its resistive channel, introducing a loss bounded
by:
1 2
E Cx ( HL ) ≤ --- ⋅ C x ⋅ V in (Eq 3-34)
2
If the dead-time is too long, the inductor discharges C x below ground, until the
NMOS Turn-Off
Ideally, the NMOS device is gated off when i Lf decays to zero. In this case, the
i Lf (t) and v x (t) waveforms will ring from the initial condition, i Lf (t) = 0, v x (t) = 0, to the
final steady-state condition during idle mode, i Lf (t) = 0, v x (t) = Vo in the resonant
circuit of Figure 3.13. Since in any practical DC-DC converter, C f » C x , in this circuit
the output capacitor is modeled as an ideal voltage source. The ringing v x (t) and i Lf (t)
The energy dissipated in this ring (in the equivalent series resistance in the L f -
1 2
E Cx ( LI ) = --- ⋅ C x ⋅ V o (Eq 3-35)
2
3.4 PFM Operation 66
iLf (t)
+ Lf +
R
vx (t) Cx Vo
- -
vx (t)
v x ( max ) = 2V o
τ LC = LfCx ( Cx « Cf )
Vo
t
iLf (t)
t
Vo
i Lf(min) = – --------------------
Lf ⁄ Cx
The LI subscript in (Eq 3-35) indicates the low-to-idle transition at v x . Note that if:
where V D is the PMOS forward bias diode voltage (approximately equal to 0.7 V), the
PMOS body diode will conduct for a portion of the first sinusoidal cycle, dissipating
additional energy.
If the NMOS device turns off too early (i Lf = I ε > 0), additional energy stored
iLf (t)
- Lf +
VD NMOS Vo
body diode
+ -
VD t
vx(t)
iLf(t)
slope = -(Vo + VD) / Lf
Id
1 2 1 2
E L = --- ⋅ L ⋅ I ε < E C = --- ⋅ C x ⋅ V D (Eq 3-37)
2 2
where V D is the forward bias NMOS diode voltage (also approximately equal to 0.7 V),
the NMOS body diode will not forward bias, and all of E L will be dissipated in the
resistance in series with the LC tank. If the condition of (Eq 3-37) is not satisfied, the
NMOS body diode will conduct, dissipating some of E L and delivering the rest to the
output. Figure 3.15 and Figure 3.16 show the equivalent circuit and i Lf (t) and v x (t)
waveforms during NMOS body diode conduction. Since the voltage drop across the
diode is large compared to that across any resistance in series with the LC tank, R is
eliminated from this model, leaving the body diode as the only dissipater. In this case,
3.4 PFM Operation 68
the ratio of energy dissipated to energy stored is equal to the ratio of voltage drop
VD
E diode = E L ⋅ -------------------------------------- (Eq 3-38)
V bat + V D – V o
where
1 2
E L = --- ⋅ L ⋅ I d (Eq 3-39)
2
and
VD
I d = I ε – -------------------- (Eq 3-40)
L f ⁄ Cx
In addition,
1 2 2
E R = --- ⋅ C x ⋅ ( V o + V D ) (Eq 3-41)
2
is dissipated in the series R before and after body diode conduction, resulting in an
1 2
energy penalty of --- ⋅ C x ⋅ V D .
2
From the above results, the total energy penalty associated with an early
1 2 VD
E penalty = --- ⋅ L ⋅ I ε I ε < -------------------- (Eq 3-42)
2 Lf ⁄ Cx
1 2 1 2 VD otherwise
E penalty = --- ⋅ C x ⋅ V D + --- ⋅ L ⋅ I d ⋅ ---------------------
2 2 Vo + V D
If the NMOS device turns off too late (i Lf = I ε < 0) some or all of the energy
stored in the inductor is dissipated in the series resistance and/or the PMOS body diode.
3.4 PFM Operation 69
Since the analysis is similar to the derivation of (Eq 3-42), only the resulting losses are
given:
1 2 1 2 VD otherwise
E penalty = --- ⋅ C x ⋅ ( V D + V bat ) + --- ⋅ L ⋅ I d ⋅ --------------------------------------
2 2 V bat + V D – V o
In (Eq 3-43),
( V D + V bat )
I d = I ε + ------------------------------ (Eq 3-44)
L f ⁄ Cx
Energy storage by the stray inductance L s in the loop formed by the input
decoupling capacitor C in and the power transistors causes dissipation (Figure 3.9). In
the PFM PMOS turn-on transition, i Lf = 0, and since no energy is stored in L s , there is
no associated loss. The PMOS turn-off / NMOS turn-on transition occurs when the peak
inductor current, I p , flowing into the power circuit is switched from the high-side to the
1 2
E Ls = --- ⋅ L s ⋅ I p (Eq 3-45)
2
The PFM control circuitry consumes static power, even when the converter is
PFM control. Since T increases with decreasing load, E static becomes the dominant
Two other basic configurations for PWM switching converters are the boost
converter (Figure 3.17) and the buck-boost converter (Figure 3.19). All three basic
topologies − buck, boost, and buck-boost − are similar in that they each have two
complementary switches and one inductor. Their conversion ratios may all be adjusted
by varying the duty cycle with frequency held constant. They can all be derived from
v x (t) waveform is shown in Figure 3.18. In one portion of the cycle, (1-D), the NMOS
device is on, and the input voltage is applied across L f , building up current and thus
storing energy in the inductor. When the NMOS switch is turned off, the attempt to
interrupt the current in the inductor causes the voltage at node v x to rise rapidly. The
Lf
+ + +
Vin Cin vx Cf Vo
- - -
PMOS
Vo
on
NMOS Vx(DC) = Vin
0 on
DTs (1-D)Ts
PMOS device is turned on at this point, limiting the voltage produced by this inductive
kick to the voltage on the output capacitor. (If the PMOS device were not turned on, its
drain-body diode would short v x to one diode drop above Vo .) During the fraction of the
cycle, D, that the PMOS device conducts, some of the energy stored in the inductor is
transferred to the output, along with additional energy flowing from the input. The
The boost converter may be considered a variation of the buck converter, but
with power flow from the lower voltage side to the higher voltage side. The voltage at
node v x is a rectangular wave whose DC component is equal to the input voltage. (It
must be equal, as the average voltage across the inductor must be zero for periodic
steady state.) Thus, the input and output voltages are related by:
V in = V o ⋅ D (Eq 3-47)
+ + -
Vin Cin vx Lf Cf Vo
- - +
PMOS
Vin
on
NMOS Vx(DC) = 0
on
Vo
DTs (1-D)Ts
the same relation as for the buck converter, but with the input and output terminals
reversed.
The operation of the buck-boost converter (Figure 3.19) is similar to that of the
buck converter, in that the cycle starts with the input voltage applied across the
inductor, in this case through the PMOS device for a duration, D ⋅ T s . However, when
the PMOS device is turned off, the voltage at v x heads downward, and the circuit
produces an output voltage polarity opposite to that of the input (Figure 3.20). The
energy transferred to C f during this portion, (1-D), of the cycle (while the NMOS
device conducts) is only the energy stored in the inductor, with none coming directly
from the input. Setting the average voltage across the inductor equal to zero allows the
D
V o = V in ⋅ ------------- (Eq 3-48)
1–D
Note that this allows input voltages of smaller or larger magnitude than the input, hence
preclude the use of a PWM DC-DC converter in many applications. Two alternatives
that do not require magnetic components are linear regulators and switched-capacitor
principle constraints. The output voltage, Vo , must be less than the input voltage, Vin ,
and the efficiency, η, can never be greater than V o ⁄ V in . However, linear regulators
have the advantage of requiring few or no reactive components, and they can be very
small and simple. This makes them especially attractive for portable applications.
voltage just slightly below the input voltage. This requirement may be incompatible
with other system design constraints, but in some systems it is practical, and, in this
Pass Device
Unregulated dc Regulated dc
Vin
+
RL Vo
-
Vref
Error Amplifier
case, a linear regulator may be highly efficient. The achievable efficiency then depends
on two parameters of the regulator: quiescent current and dropout voltage. The
quiescent current determines the regulator’s dissipation when the load is not drawing
regulation is lost, and the output voltage will sag below the nominal regulation point.
Dropout voltage is this minimum voltage difference between input and output required
to maintain regulation. If it is not very low, it can conflict directly with the design
requirement of having the output voltage only slightly less than the input voltage, and
will therefore preclude high efficiency. This becomes especially important in low-
increase in the minimum input power over what would be required with zero dropout
voltage. However, with a 1 V output, a 1 V dropout voltage doubles the minimum input
power.
Linear regulator circuits with low quiescent power, and PNP or MOSFET pass
devices to allow low dropout voltage, are now commercially available. In the limited
class of circuits that require a regulated voltage just below the input voltage of the
in ICs where a voltage higher than, or of opposite polarity to, the input voltage is
applications are usually limited to those in which poor efficiency and very low output
φ1 φ2
+ +
φ2 Cs
Vin Vo
φ1
- -
Fig. 3.22: A switched-capacitor voltage doubler. Switches labeled φ1 and φ2 are closed alternately.
voltage doubler. The switches are closed in pairs, alternately. First the switches labeled
φ 1 are closed, charging capacitor C s to the input voltage, Vin . Then the φ 1 switches are
opened, and the φ 2 switches are closed. This places C s , which is now charged to Vin , in
series with the input voltage, producing a voltage of 2 ⋅ V in across the output. The cycle
then repeats. The output capacitor maintains the output voltage near 2 ⋅ V in during φ 1 .
The same converter topology can be used as a step-down converter, producing an output
voltage of half the input voltage, by exchanging the input and output terminals. By
ratio, for example by first stepping the voltage up by one integer ratio, and then
stepping down by another integer ratio. Some of the many possible topologies are
entirely of theoretically lossless elements − in this case, only switches and capacitors.
resistances in the capacitors and switches approach zero, the loss in the converter
approaches a non-zero limit. This is in contrast to a PWM converter, in which the losses
dissipation which occurs when a pair of capacitors, charged to different voltages, are
shorted together through a switch. If two capacitors with values C1 and C 2 , initially
charged to voltages V 1(initial) and V 2(initial) , respectively, are shorted together through a
1 C 1 C2 2
E diss = --- ⋅ ------------------- ⋅ ( V 1 ( initial ) – V 2 ( initial ) ) (Eq 3-49)
2 C 1 + C2
doubler shown in Figure 3.22. During φ 2 , the equivalent circuit is as shown in Figure
3.23. The charge flowing to the output is supplied by both the input and C s . During φ 1 ,
this same quantity of charge must be supplied from the input and stored on C s for the
next cycle. Since all the charge that flows out of the output must be supplied twice by
the input, the average input current must equal twice the average output current, i.e.,
Vo ⋅ Io Vo
η = ------------------- = ---------------- (Eq 3-50)
V in ⋅ I in 2 ⋅ V in
+
Cs
+ Vo
Vin
- -
The efficiency would be 100% if Vo were in fact twice Vin . However, in order
for a charge, Q, to flow into C s during φ 1 and subsequently flow out of C s during φ 2 , the
voltages applied across C s during the two phases must differ by an amount
resistance of the switches and C s is small compared to the switching period so that the
charge on C s reaches its steady-state value before the end of each phase, and that the
input and output capacitors are large enough to maintain constant Vin and Vo , the
Io ⋅ Ts
2 ⋅ V in – V o = -------------- (Eq 3-51)
Cs
The circuit may be modeled as shown in Figure 3.24, with an ideal doubler
that accounts for the voltage drop ∆V. The effective resistance also accounts for the
loss; calculating the dissipation in this resistor gives a result identical to that found
transformer with a fixed rational turns ratio, N, and an effective resistance. The
conversion ratio, N, can be chosen to bring Vo near the desired output voltage; to
Reff = Ts / Cs
1:2
+ +
Vin Vo
- -
precisely regulate Vo , R eff is varied through changes in the switching frequency. Using
R eff for regulation is undesirable, since increasing it to lower the output voltage
produces additional power dissipation. However, N is fixed by the topology, and cannot
efficiently convert voltages, but they cannot regulate these converted voltages any more
but regulation is not necessary, or to situations in which the regulation range is limited,
hold, it is necessary for the time constant of the switched capacitor and the on-
resistance of the switch to be much less than the switching period, i.e. C s ⋅ R on « T s .
This requires the use of a large MOSFET to implement the switch, but the gate-drive for
that device then requires substantial power, especially if a high switching frequency is
used to minimize the required size of C s . Thus, gate-drive loss must be considered in
the design.
of its plates to ground will be a substantial fraction of its terminal capacitance. This
2
introduces C stray ⋅ V ⋅ f s loss, further hampering efficiency. Technologies for
fabricating capacitors with low stray capacitance to ground, or off-chip capacitors are
Chapter 4
were listed in Chapter 3. In this chapter, design techniques at the power system,
individual control system, and circuit levels are described which help meet the stringent
applications, where small size and high efficiency are the principal objectives.
Section 4.2, circuit and control system techniques are described which eliminate,
minimize, or reduce the power dissipation due to each primary loss mechanism. Several
and mass, the volume and mass of a converter can be a critical design consideration.
This section introduces several design techniques that may be used to reduce both the
As indicated by (Eq 3-2) and (Eq 3-3), there are inherent size and cost
advantages associated with higher frequency operation. The reactive filter components
f s = ω s ⁄ ( 2π ) , enables the use of reactive components with smaller value and smaller
–1
physical size. Ideally, the size of these components will decrease with f s . However, as
will be described in Section 4.2.4, if the operating frequency of the circuit is increased,
the sum of the losses in the power transistors and drive, if optimized, will increase
roughly with f s . Thus, the general theoretical relationship between the size of a DC-
DC converter and its losses is as illustrated in Figure 4.1. Here, operating frequency is
used as a parameter, and the sum of the losses in the power transistors and drive is
If the cost and volume of the converter are decreased, additional space and
resources are left for a larger or better battery, compensating for lower conversion
3
Normalized Losses
Increasing fs
2
Fig. 4.1: General trends in power transistor losses versus the size of a DC-DC converter.
4.1 Converter Miniaturization 81
efficiency. The system requirements and battery characteristics will help to determine
which point on this curve is optimal for a specific application. For example, in systems
designed for shorter run-times, the volume of the converter can become comparable to
the volume of the battery, particularly if a battery with a relatively high volumetric
energy density is used. Then, it might be worthwhile to operate the converter at a higher
frequency, sacrificing efficiency while leaving space for additional battery capacity.
reduce the frequency-dependent losses in the power train, yielding a class of miniature
yet highly efficient converters that are well-suited for portable applications. In practice,
power train and controller, but also by diminishing returns in the miniaturization of the
Since the L f C f product determines the output voltage ripple (Eq 3-3), the
relative size and cost of inductance versus capacitance should be considered in the
selection of these components. As the size, cost, and commercial availability of low-
voltage multilayer ceramic chip capacitors are often superior to those of inductors,
using large-value capacitors and small-value and small-size inductors is preferred. This
decision is restricted primarily by the increasing rms current in the inductor, which
2
i Lf ( rms ) .
to-peak ripple ∆I superimposed on the DC output current, I o , (see Figure 3.7). In Figure
4.2, ∆I is varied, and its effects on three key circuit parameters are shown. As
4.1 Converter Miniaturization 82
10
(Conduction Loss)
(Physical Size)
E = 1/2 Lf (Io + ∆I/2)2
0.1
Lf ∝ ∆I-1
10-20.1 1 10
Current Ripple ∆I (Normalized to Load Io)
Fig. 4.2: The effect of increased current ripple on the value of Lf, the physical size of Lf, and
iLf(rms)2.
illustrated by (Eq 3-2), the value of filter inductance decreases with ∆I -1 . However, the
physical size of L f is roughly proportional to its peak energy storage, which in turn, is
given by:
∆I 2
E Lf = --- ⋅ L f ⋅ I o + ------
1
(Eq 4-1)
2 2
∆I 2
I o + --- ⋅ ------
2 1
i Lf ( rms ) = (Eq 4-2)
3 2
and for ∆I = 2I o , the AC component of the current accounts for 25% of the overall full-
between size and loss in a particular application, it can be concluded that a peak-to-
peak current ripple in the range I o < ∆I < 2Io is optimal for many applications. As ∆I is
4.1 Converter Miniaturization 83
in conduction loss at full load, and do not justify the larger inductor that would be
required. There is no obvious benefit for ∆I > 2I o , but this will be seen to be
A completely monolithic supply (active and passive elements) would meet the
severe size and weight restrictions of a hand-held device. Because most portable
applications call for low-voltage power transistors, their integration in a standard logic
inductors of suitable value and quality for efficient power conversion [Barringer93].
[Sullivan93]), but currently, magnetics, capacitors, and silicon circuitry are fabricated
separately and assembled at the board level or in a multi-chip module (MCM). The
extent of integration is the use of a monolithic silicon circuit, including all power
Such a highly-integrated solution not only results in a more compact and cost-
effective design, it gives the designer more latitude in physical design and device
efficient converter. Parasitics from both the active devices and interconnect may be
energy storage of these parasitics; thus, integration enables higher efficiency at high
integrating multiple supplies on a single die, and integrating small custom DC-DC
converters with their individual loads, the overall size and cost of the entire power
converter have been summarized in Section 3.3.2 and Section 3.4.2. In this section,
mechanisms are described. While the following discussion is sometimes specific to the
buck circuit, all of the techniques presented here can be applied to maximize the
efficiency of boost and buck-boost type converters, each of which is typically required
The focus of this chapter is the CMOS low-voltage buck converter, in which
the switching elements, modeled by the single-throw double-pole switch in the block
Consider the conventional buck circuit of Figure 4.3. Even if all other losses in
the circuit are made negligible, the maximum efficiency is limited by the forward bias
diode voltage, V diode . Since the diode conducts for a fraction (1-D) of the switching
period, the maximum efficiency this circuit can obtain is given by:
Vo
η max = ---------------------------------------------------- (Eq 4-3)
V o + ( 1 – D ) ⋅ V diode
For example, consider a conventional buck circuit used to generate an output voltage of
1.5 V from a single lithium ion cell. Even using a low-voltage Shottky diode with a
forward drop of 0.3 V, at the nominal cell voltage of V in = 3.6 V , η max is lower than
90%. With a silicon bipolar diode, V diode = 0.7 V , and η max = 0.79 .
If the diode in Figure 4.3 is replaced by an NMOS device which is gated when
the diode would have conducted (M n in Figure 3.2), the forward drop can be made
arbitrarily small by making the device sufficiently large. In this way, the NMOS device,
used as a synchronous rectifier, can perform the same function as the diode more
efficiently. Assuming all other losses, including the gate-drive for the synchronous
rectifier, are still negligible, the maximum efficiency of the low-voltage buck converter
approaches unity.
Although the synchronous rectifier may reduce conduction loss at low output
voltage levels, it comes at the expense of an additional gate-drive signal and its
+ S1 Lf +
Vin Cin D1 Cf Vo
- -
Fig. 4.3: Conventional buck circuit with pass device, S1, and diode.
4.2 Circuit Techniques for High Efficiency 86
associated loss. In addition, as mentioned in Section 3.3.2 and Section 3.4.2, without
proper control of the rectifier, a short-circuit path may exist temporarily between the
input rails during transients. In the rectifier control scheme described in Section 4.2.3,
the dead-times, which ensure that M p and M n never conduct simultaneously, are
adjusted in a negative feedback loop to achieve nearly ideal zero-voltage switched turn-
2
power in proportion to C x ⋅ V in ⋅ f s as a result of the step charging of parasitic
substantial short-circuit loss (if no dead-time is provided), or reverse recovery loss (if a
source to charge and discharge this capacitance in an ideally lossless manner, allowing
transitions. In this way, appropriate dead-times may be set such that the power
loss.
4.2 Circuit Techniques for High Efficiency 87
+
Mp iLf
vx
Vin Cin
Lf +
Mn Cx Cf Vo
- -
Ts
vx
DTs Vin
τxLH τxHL t
iLf
∆I/2
Io
∆I/2
t
Mp ON
|Vgsp|
Vgsn
t
Mn ON
Figure 4.4 and Figure 4.5 show the low-voltage buck circuit and associated
periodic steady-state waveforms for ideal zero-voltage switching operation. The soft-
Assume that at a given time (the origin in Figure 4.5), the rectifier M n is on, shorting
the inverter output node to ground. Since by design, the output is DC and greater than
the value of filter inductance is small enough, the zero-to-peak current ripple exceeds
4.2 Circuit Techniques for High Efficiency 88
the full load ∆I > 2I o , and i Lf ripples below zero. As illustrated in Section 4.1.2, for ∆I
slightly larger than 2I o , the physical size of the inductor is close to minimum.
If the rectifier is turned off after the current reverses (and the PMOS device,
M p , remains off), L f acts approximately as a current source, charging the inverter output
node. To achieve a lossless low-to-high transition at the inverter output node, the
PMOS device is turned on when v x = V in . In this scheme, a pass device gate transition
With the PMOS device on, the inverter output node is shorted to Vin . Thus, a
constant positive voltage is applied across L f , and i Lf linearly increases, until the high-
time, the sign of current i Lf is positive. Again, L f acts as a current source, this time
and discharge all capacitance at the inverter output node (and supply all Miller charge)
transitions. Since the power transistors are switched at zero drain-source potential, this
capacitive switching loss. Furthermore, because the inductor current in a ZVS circuit
reverses, if the body diode conducts for a portion of the cycle, it turns off through a
short circuit (rather than through a potential change of Vin ), nearly eliminating the
dissipation associated with reverse recovery, a factor which might otherwise dominate
To ensure ideal ZVS of the power transistors, the periods when neither
conducts (the dead-times), τ D , must exactly equal the inverter output node transition
times:
inductor current ripple is symmetric about the average load current. As the average load
varies, the DC component of the i Lf waveform is shifted, and the current available for
commutating the inverter output node is modified. Thus, the inverter output node
yielding estimates of the inverter output node transition times. Fixed dead-times are
based on these estimates. In this way, losses are reduced, yet perhaps not to negligible
levels.
to soft-switching may not be adequate. To illustrate the potential hazards of fixed dead-
time operation, Figure 4.6 shows the impact of non-ideal ZVS on conversion efficiency
through reference to a high-to-low transition at the inverter output node. In Figure 4.6a,
the dead-time is too short, causing the NMOS device to turn on with v dsn > 0 , partially
discharging C x through a resistive path and introducing losses. Since shunt capacitance
with a value much larger than the intrinsic parasitics may be added to slow the soft-
switched transitions in a ZVS circuit, this loss may be substantial. In Figure 4.6b, the
dead-time is too long, and the inverter output node continues to fall below zero until the
4.2 Circuit Techniques for High Efficiency 90
vx τxHL vx τxHL
Body diode
conduction
Mn discharges Cx
(a) (b)
forward-bias body diode voltage is a significant fraction of the output voltage; thus,
body diode conduction must be avoided for efficient operation. When the rectifier (M n )
turns on, it removes the excess minority carrier charge from the body diode and charges
control scheme for a 1 MHz ZVS buck circuit has been outlined in [Stratakos94]. Figure
4.7 shows a block diagram of the approach. A phase detector updates an error signal
based on the relative timing of v x and the gate-drive signals of the power transistors. A
delay generator adjusts the dead-times based on these error signals. Using this
technique, effective ZVS is ensured over a wide range of operating conditions and
was made in [Acker95], and a successful IC implementation of a ZVS buck circuit was
reported in [Lau97].
pump. An error voltage proportional to the difference between the high-to-low soft-
switched inverter output node transition time and its corresponding dead-time is
4.2 Circuit Techniques for High Efficiency 91
vgn
vx phase delay
detectors generators
error vgp
signals
vgn
vgp
τDLH τDHL
generated on integrating capacitor, C I . This error voltage is sampled and held at the
Vin
V in ( nT s ) – v ε ( nT s )
i control ( nT s ) ≈ ------------------------------------------------ (Eq 4-7)
R
C ⋅ V M+
τ DHL ( nT s ) ≈ ------------------------------------------ (Eq 4-8)
i control ( nT s – T s )
In periodic steady-state, the error voltage, and thus the gate timing errors, are
forced to zero, nulling propagation delays in the control and drive circuitry. Figure 4.9
shows the periodic steady-state waveforms associated with an ideal ZVS rectifier turn-
on.
vgp
vx Vin / 2
vgn
2I
iCI
0
-I
τxHL / 2
τDHL
Fig. 4.9: Ideal steady-state waveforms for the τDHL adjustment loop.
4.2 Circuit Techniques for High Efficiency 93
A similar loop is used to adjust the dead-time between the turn-off of M n and
essentially eliminated. If the filter components in the buck circuit of Figure 4.4 are
ideal, and series resistance and stray inductance in the power train are made negligible,
the fundamental mechanisms of power dissipation will include on-state conduction loss
and gate-drive loss in the power transistors. When sizing a MOSFET for a particular
power application, the principal objective is to minimize the sum of the dissipation due
to these mechanisms. This minimization is performed at the operating point where high
efficiency is most critical: Usually at full load, at high temperature, and in portable
–1
the triode region, where r ds = R 0 ⋅ W (the channel resistance is inversely
2
i ds ( rms ) ⋅ R 0
P q = ----------------------------- (Eq 4-9)
W
Since the device parasitics generally increase linearly with increasing gate-
P g = E g0 ⋅ f s ⋅ W (Eq 4-10)
transition cycle (see Section 4.2.5 for more detail) and f s is the switching frequency of
4.2 Circuit Techniques for High Efficiency 94
the converter. In a ZVS circuit, the filter inductor supplies all of the Miller charge, so
2
i ds ( rms ) ⋅ R 0
W opt = ----------------------------- (Eq 4-11)
E g0 ⋅ f s
2
P q ( opt ) = P g ( opt ) = i ds ( rms ) ⋅ R 0 ⋅ E g0 ⋅ f s (Eq 4-12)
and P total = P q + P g is at its minimum value, P t(min) . Figure 4.10 illustrates normalized
Pt = Pg + Pq
Pq = i2ds-rmsRo/W
Pg = EgofsW
0
0 1 2
Gate-Width (Normalized to Wopt)
x1 x1 x1
CLK
CLK
CLK
CLK xN
φ0 φ1 φN
DYNAMIC FET SIZING DECODER
A/D CONVERTER
LOAD Vin
Note that dynamic power transistor sizing may be used to repeat this
optimization at various battery voltages, as terms i ds(rms), R 0 , Q g0, and Vin , are each a
function of battery voltage. Since the battery voltage is slowly varying with time, a
slow, low-precision A/D might be used to quantize the battery voltage every few
reasonably short in a ZVS buck circuit, the inductor current ripple must be made
substantial. This gives rise to large circulating currents in the power train, and
therefore, when the power transistors are sized according to (Eq 4-11), increased gate-
drive losses. Since gate-drive losses increase in direct proportion with f s , this proves to
demonstrated the ability to recover a significant fraction of the gate energy at lower
frequencies, due to the resistance of the polysilicon gate of a power transistor, none are
applications.
approach to reducing gate-drive dissipation is to reduce the gate energy consumed per
systems where a number of low-voltage supplies are typically available for the gate-
drive. However, because the channel resistance of the device increases with
–1
( V g – Vt ) , gate-swing cannot be arbitrarily reduced, implying the existence of an
optimum V g .
of value C g over the voltage range 0 ≤ V GS ≤ V g , the gate energy dissipation in a single
2
Eg = Cg ⋅ V g (Eq 4-13)
Since the power transistors conduct almost exclusively in the triode region, where:
∂I D W
g ds = -------------- ≈ µC ox ----- ( V g – V t ) (Eq 4-14)
∂V DS L
1
R DS = ------- (Eq 4-15)
g ds
according to (Eq 4-11), its total dissipation is minimized, and that this minimum
dissipation is related to the square root of the product of the gate energy and the device
channel resistance:
Vg
P t ( min ) ∝ R 0 ⋅ E g0 ∝ ----------------------- (Eq 4-16)
Vg – V t
V g = 2V t (Eq 4-17)
While the total dissipation of a power transistor may be reduced by lowering V g (for
V g > 2V t ) and appropriately scaling its gate-width, the optimum gate-width which
technology dependent. For V g < V t , the channel of the device is not enhanced, and the
incremental gate capacitance may, to the first order [Rabaey96], be approximated by the
Wopt
01 3 5 7
Gate-Swing Vg (Normalized to Vt)
Fig. 4.12: The optimal gate-width and minimum total dissipation for a power MOSFET versus
gate-swing in a ZVS topology.
Gate Capacitance, Cg
triode: WLCox
cut-off:
2WLDCox
0
Vt
On-State Gate-Source Voltage Vg
Fig. 4.13: A first-order gate capacitance model to a power MOSFET in a ZVS application.
dQ g
C g = ---------- ≈ 2WL D C ox, ( Vg < Vt ) (Eq 4-18)
dv gs
where L D is the lateral diffusion in the drain and source areas, and C ox = ε ox ⁄ t ox is the
gate oxide capacitance per unit area. For V g > 2V t , the channel is enhanced, and
4.2 Circuit Techniques for High Efficiency 99
dQ g
C g = ---------- ≈ WLC ox, ( Vg ≥ Vt ) (Eq 4-19)
dv gs
Here, L is the drawn channel length, and is equal to the sum of the effective channel
length and the lateral diffusion in both the source and drain diffusion areas (see Figure
4.14):
Note that in a ZVS circuit, the Miller charge is supplied by the filter inductor through
the drain, not through the gate-drive. Thus, the effective gate capacitance does not
Polysilicon Gate
Source Drain
Leff W
n + LD LD n+
Gate-Bulk
L Overlap
(a)
Gate Oxide
tox
n+ Leff n+
(b)
Fig. 4.14: An illustration of the effect of lateral diffusion, LD, on the effective channel length, Leff,
of a power MOSFET: (a) Top view. (b) Cross-section.
4.2 Circuit Techniques for High Efficiency 100
E g , is:
E g = V g ⋅ ∆Q g (Eq 4-21)
where V g is the potential of the gate-drive supply voltage, and ∆Q g is the change in
t t Vg
C g dv g’
∆Q g = ∫ ig dt’ = ∫ ---------------
dt’
- dt’ = ∫ Cg dv g’
0 0 0
Vt Vg
(Eq 4-22)
= 2WL D C ox V t + WLC ox ( V g – V t )
Thus, neglecting dissipation due to the inverter chain, the total gate energy dissipation
2
E g = WLC ox V g – WL eff C ox V g V t (Eq 4-23)
Substituting (Eq 4-23) into the expression for P t(min) in (Eq 4-12), and
minimizing this total dissipation with respect to V g , the optimum gate-drive voltage is:
2L
V g ( opt ) = V t ⋅ 1 + ---------D- (Eq 4-24)
L
which is process technology dependent and less than 2V t . For a standard 1.2 µm digital
CMOS process in which L D ≈ 0.15 µm , (Eq 4-24) yields V g ( opt ) = 1.5V t , or about 1.2 V
decreased below 2Vt , this component begins to dominate the overall gate-drive
While (Eq 4-25) is useful for first-order design centering, iteration with a circuit
4.2.5.3 Scaling Vt
To further reduce the total dissipation of a power MOSFET with a given gate
voltage swing, the off-state voltage can be made greater than zero (Figure 4.15a) to
increase the gate overdrive, reducing the device channel resistance. This scheme is
equivalent to that shown in Figure 4.15b, where V GS = 0 in the off-state, and the
device threshold voltage, V t' < V t , is scaled, while all other parameters are held
constant, if:
(ON)
(a) (b)
Fig. 4.15: Two equivalent schemes to further reduce total power transistor losses: (a) The gate-
source voltage is not brought to zero. (b) Lower Vt.
4.2 Circuit Techniques for High Efficiency 102
power dissipation in the power train of the converter. Figure 4.16 shows the inherent
compromise associated with V t scaling. Here, using the simple zero-order model for
gate energy consumption and the model for subthreshold current conduction presented
in [Liu93], the optimal gate-width and minimum total dissipation of an NMOS power
room temperature and with all other application- and technology-related parameters
held constant. The gate-swing has been optimized for minimum dissipation ( V g = 2V t ) ,
and subthreshold conduction has been considered in the selection of optimum gate-
width. For V t > 0.4 V , leakage power dissipation (at V in = 6 V ) is negligible compared
to the gate-drive power (at f s = 1 MHz ), and as Wopt increases with 1 ⁄ V t , P t(min)
decreases with V t . As the threshold voltage is dropped below 0.4 V, leakage power
inverters which are scaled with a constant tapering factor, u, such that
N Cg
u = ------ (Eq 4-27)
Ci
4.2 Circuit Techniques for High Efficiency 103
1.8
1.2 µm CMOS
1.2 Pt-min
0.8
Wopt
Fig. 4.16: The optimal gate-width and total minimum dissipation, including static power
dissipation due to subthreshold conduction, for a power n-channel MOSFET versus Vt in a 1 MHz
ZVS buck circuit.
Here, C g is the gate capacitance of the power transistor and C i is the input capacitance
of the first buffering stage. This scheme, depicted in Figure 4.17, is designed such that
the ratio of average dynamic current to load capacitance is equal for each inverter in the
chain. Thus, the delay of each stage and the rise/fall time at each node are identical. It
is a well known result that under some simplifying assumptions, the tapering factor u
that produces the minimum propagation delay is the constant e [Mead80]. However, in
power circuits, the chief concern lies not in the propagation delay of the gate-drive
Vg Vg Vg 0 ↔ Vg
τ x » τ gs ≈ uτ 0 (Eq 4-28)
where τ x is the soft-switched inverter output node transition time, τ gs is the maximum
gate transition time which ensures effective ZVS of the power transistor, τ 0 is the
output transition time (rise/fall time) of a minimal inverter driving an identical gate,
and u is the tapering factor between successive inverters in the chain. In general, it is
desirable to make τ gs as large as possible (yet still a factor of five to ten less than τ x ),
minimizing gate-drive dissipation. Given τ gs and τ 0 , if there exists some u > e such that
the criterion given by (Eq 4-28) is met, the buffering scheme of Figure 4.17 will
provide a more energy efficient CMOS gate-drive than that obtained through
minimization of delay.
In this analysis, a minimal CMOS inverter has an NMOS device with minimum
the NMOS device. It has lumped capacitances C i at its input and C o at its output. Given
that the pull-down device operates exclusively in the triode region during the interval
the output fall time of a minimal inverter driving an identical gate from
Co + C i
τ 0 = ------------------ ⋅ κ (Eq 4-29)
W0
which is linearly proportional to the capacitive load, inversely proportional to the gate-
width of the n-channel device, and directly related to the application and technology
dependent constant:
4.2 Circuit Techniques for High Efficiency 105
2L ( 2V g – 3V tn ) ( V g – V tp )
κ ≡ ------------------------------------------ ⋅ log --------------------------------- ⋅ ----------------------------------------------- (Eq 4-30)
µ n C ox ( V g – V tn ) V tn ( V g – 2V tn + V tp )
In [Chandrakasan94b], a similar expression can be found for the output fall time
solving:
κ ( C o + uC i )
τ gs = ------------------------------ ≈ uτ 0 (Eq 4-31)
W0
τ gs W 0 – κC o
u = -------------------------------- (Eq 4-32)
κC i
between successive buffers. Given u, the number of inverters in the chain is:
log ( C g ⁄ C i )
N = ----------------------------- (Eq 4-33)
log ( u )
Loss Analysis
2
P dyn = C T ⋅ V g ⋅ f s (Eq 4-35)
where I sc, i is the mean short-circuit current in the i th inverter in the chain, and the total
switching capacitance, including the loading gate capacitance of the power MOSFET, is
N
) ⋅ ( C o + C i ) + C g = --------------- ⋅ ( C o + C i ) + C g
2 N–1 u –1
CT = ( 1 + u + u + … + u (Eq 4-37)
u–1
Since u N is the constant given by (Eq 4-27), C T and thus, the dynamic dissipation, is
Though the dynamic component is readily calculated from (Eq 4-35) and (Eq
4-37), the short-circuit dissipation is more difficult to quantify. From Figure 4.18, it can
be seen that short-circuit current exists in a CMOS inverter while the n- and p-channel
devices conduct simultaneously ( V tn < V in < V g – V tp ) , and that the total energy
consumed during an input transient is proportional to both the input transition time and
Vg - |Vtp|
Inverter input
Vtn
Ipeak
Short-circuit current
τr τf
the peak short-circuit current (which in turn, is related to the output transition time
dynamic dissipation per cycle versus the ratio of 10%-90% input to output transition
transition time, but is negligible for equal input and output signal transition
(I peak increases linearly with device size), the ratio of E sc to E dyn appears to be
independent of size.
10
1.2 µm CMOS
1
0.1
Esc / Edyn
10-4
1 3 5
τin / τout
Fig. 4.19: Simulation results showing normalized short-circuit energy versus normalized 10%-
90% input edge rate for CMOS inverters in a 1.2 µm CMOS technology.
4.2 Circuit Techniques for High Efficiency 108
While the 10%-90% input edge rate is relatively independent of supply voltage
zero.
chain, providing equal transition times τ gs at each node, the short-circuit dissipation is
made negligible, particularly at low supply voltages. Furthermore, for u > e , less silicon
area will be devoted to the buffering; thus parasitics, and ultimately, dynamic energy
on-to-off gate transition cycle of a minimal power MOSFET, (Eq 4-35) and (Eq 4-37)
are used in conjunction with the values of u and N derived in (Eq 4-32) and (Eq 4-33),
giving:
2 κ ( Co + C i )
E g0 ≈ C g0 ⋅ V g ⋅ ------------------------------------------------- + 1 (Eq 4-38)
τ gs W 0 – κ ( C o + C i )
dissipation in the inverter chain is negligible compared to the dynamic dissipation, that
N
all capacitances scale linearly with gate-width, and that u » 1 . Under these
Scaling Vg
delays in the drive circuitry, which produce reversing returns in the reduction of gate
4.2 Circuit Techniques for High Efficiency 109
This phenomenon is illustrated in Figure 4.20, where the output signal rise and
versus the supply voltage, V g . For V g > 3 V , delays are indeed relatively independent of
supply voltage, and the rise and fall times are nearly equal. However, as the supply is
the output rise time increases more quickly than the output fall time. To achieve
balanced rise and fall times at the output of a CMOS inverter with a supply voltage
5
10%-90% Transition Time [ns]
Rise Time
4
3 1.2 µm CMOS
Wn = 2.4 µm
Wp = 7.2 µm
2 Ln = Lp = 0.9 µm
Fall Time
0
1 2 3 4 5
Vg [V]
Fig. 4.20: Simulated output rise and fall times for a minimal CMOS inverter driving an identical
gate as a function of supply voltage, Vg.
4.2 Circuit Techniques for High Efficiency 110
Figure 4.21 plots the total gate energy consumed per cycle as a function of the
gate-drive supply voltage. Here, power transistor size and τ rise = τ fall = τ gs = 5 ns (a
typical gate transition time for a 1 MHz ZVS power circuit) are held constant. For
voltage. However, because of the increase in inverter output signal transition times, and
the increase in buffer input and output capacitances associated with larger p-channel
successive inverters in the chain becomes small, and the dynamic energy consumed by
the gate-drive buffering increases dramatically and begins to dominate over that
required by the gate capacitance of the power transistor. Thus, when the dissipation in
15
1.2 µm CMOS
Vtn = 0.7 V
|Vtp| = 0.9 V
10
Normalized Eg
increasing u
0
1 2 3 4 5
Vg [V]
Fig. 4.21: Gate energy per cycle (including the CMOS drive) versus gate-drive supply voltage
for fixed power transistor size and τgs = 5 ns.
4.2 Circuit Techniques for High Efficiency 111
4.2.5.5 Optimum Vg
available for the gate-drive. While this low sub-system operating voltage may not be
optimal, it is likely to be useful to reduce the minimum achievable FET losses in the
power train of each DC-DC converter. Thus, it is important to compare the minimum
achievable FET losses and the gate-width required to achieve this minimum loss for
In Figure 4.22, Wopt and P t(min) are plotted versus V g . Simulation results on a
large area n-channel MOSFET in a 1.2 µm CMOS technology have been interpolated to
find device parameters R 0 and E g0 at each data point. Dissipation in the drive circuitry
is included in E g0 . From this plot, it can be seen that the greatest power savings with
–3 ⁄ 2
( W opt ∝ R 0 ⁄ E g0 ∝ V g ) , as indicated by (Eq 4-16), P t(min) decreases as Vg .
However, since both R 0 and E g0 increase as V g is brought below the sum of the
threshold voltages in the gate-drive buffers, P t(min) increases with any further decrease
source. From Figure 4.22, the total losses in each power FET at V g = 1.5 V (the
operating voltage for the baseband circuitry in the current InfoPad terminal) are 20%
lower than at V g = 3.6 V (the nominal battery source voltage). However, the gate-
4.2 Circuit Techniques for High Efficiency 112
1.5
Pt-min
Vg ≈ Vtn + |Vtp|
0
1 2 3 4 5
Gate-Swing Vg [V]
Fig. 4.22: The optimal gate-width and minimum total dissipation for a power NMOS versus
gate-swing in a 1.2 µm CMOS technology.
width of each device must be increased by a factor greater than 4.7 to achieve this
reduced dissipation.
circuit. The gate of M n is actively driven from 0 to V g by its CMOS gate-drive. The
with an AC-coupled gate-drive. PMOS device M off , whose gate swings from rail-to-rail,
provides a low-impedance path from the gate of M p to Vin , ensuring that M p remains
Vin
φoff
Moff Vin
Vg Vg
φp
Mp
Vgp
Cc iLf
Io
Vx
Vg Vg Lf
φn Vgn
Mn Cx Cf RL
Fig. 4.23: A reduced gate-swing CMOS buck circuit implementation with gate supply Vg.
Vg
φP
0
Vin
φoff Mp conducts
0
Vin
C
c
Vgp V ⋅ -----------------------
g C + C
c gp
Vg
φN
0
Vg
Vgn 0 Mn conducts
Dead-times Dead-times
2N - 1 reference
ramp
reference
N-bit ramp control
counter N
^ 0
To
converter
N bits To
2N fs oscillator
≥
converter
control N
+
N-bit
+/- (D, fs)
N
counter
Digital
increment /
Filtering
decrement
Vref Vo
From
converter
and converted to a one-bit digital signal through a slicer with switching threshold Vref .
bit duty cycle control signal consists of this integral term, and a proportional term
stability.
also clocked at 2 N f s , compares the reference ramp and the control signal, generating a
control
D = ----------------- (Eq 4-40)
N
2 –1
the operating voltage (typically, the lowest voltage available to the system may be
–N
∆D = 2 ⋅ Ts (Eq 4-41)
is also reduced. This may result in a larger low-frequency output voltage ripple due to
Nearly any load in a portable electronic system can vary by several orders of
magnitude during system run-time. Since these loads spend a majority of time idling in
a low-power state, it is the overall energy efficiency, rather than the peak power
efficiency of the converters supplying these loads, which ultimately determines the
battery run-time.
Figure 4.26 shows the losses versus load of a high-efficiency low-voltage DC-
DC converter under three modes of control: PWM operation (see Section 3.3); PFM
operation (see Section 3.4); and a hybrid PWM-PFM control scheme. In the hybrid
4.3 System-Level Considerations 116
PWM, PFM, and Hybrid PWM−PFM Converter Losses vs. Output Power
0.09
0.08
0.07 PWM
PFM
PWM−PFM
0.06
0.04
0.03
0.02
0.01
0
−3 −2 −1 0
10 10 10 10
Pout [normalized to full load]
scheme, the converter automatically selects its control mode for peak efficiency as a
[Wang97], an adaptive hybrid control scheme is proposed which does not require
current sensing. Perhaps the most straightforward implementation ties the converter’s
operation mode directly to the pulse timing of the load. A feedforward command from
the load signals a transition from idle to full operation and vice-versa.
By considering the battery source and all DC-DC converters as a unified power
and reduce the physical size and cost of each individual DC-DC converter. This section
In the design of a complete power delivery system, the size and efficiency of
different converters within the system may be traded, and the relative merits of
different topologies may be considered in the selection of the battery source voltage.
Furthermore, resources such as oscillators and reference voltages may be shared among
components on the same die, and various sub-system voltages may be utilized in the
incorporated in the overall design, the resulting power system is likely to be far
chosen to minimize component count. The three basic topologies described in Chapter
3, buck, boost, and buck-boost, each require two switches, two capacitors, and one
inductor − the minimum component count for a PWM DC-DC converter. However, they
are a small subset of the many DC-DC converter topologies that have been proposed
and that are used in practice. Other important classes of converter topologies include
below), and soft-switching can be achieved without the use of resonant techniques
(Section 4.2.2). Thus the basic topologies are appropriate, perhaps optimal, for most
directly from the input to the output, reducing the energy storage requirement of the
inductor, and thus, its physical size. In a buck-boost converter, because none of the
energy is transferred directly − it is transferred from the input into the inductor, and
then in a separate portion of the cycle, from the inductor to the output − a larger
inductor is typically needed in this circuit. Thus, the buck and boost topologies are
topology should only be used for voltage polarity inversion, or in applications which
require both up-conversion and down-conversion over the discharge of the battery
source.
that they require no external magnetic components, were introduced in Section 3.6.
There it was shown that their efficiency is fundamentally limited by the conversion
ratio. They should therefore be used judiciously in applications where physical size and
cost are of far greater concern than energy dissipation, or where the conversion ratio
(over the entire battery discharge) is within a range that allows an acceptable energy
efficiency.
desirable to accomplish conversion over a wide voltage ratio, because the turns ratio in
the transformer can produce most of the voltage ratio. This allows switching patterns
similar to those in a 1:1 converter (see below), minimizing inductor requirements (and
outweigh any size reductions that would result from decreased inductor requirements.
Thus, transformer-coupled circuits are likely to be useful in portable systems only for
special applications, and will not be discussed further. Special applications that could
4.3 System-Level Considerations 119
requirements (e.g., for a display or backlight) and isolation. The reader is referred to
[Kassakian91], and the references contained therein, for more details on these circuits.
constraints may dictate the selection of battery voltage and converter output voltage
and thus determine the required conversion ratio, in the design of a complete power
inductor size. For example, in (Eq 3-2), it is shown that for a PWM buck converter with
a given output voltage, the required inductor value is proportional to the complement of
the duty cycle, (1-D). Thus, as the conversion ratio approaches 1:1, D approaches one,
and the value and physical size of the inductor approach zero. Similarly, the inductor
1:1. In a buck-boost converter, a 1:1 ratio still minimizes the inductor requirement, but
the requirement does not approach zero as the conversion ratio approaches 1:1.
possible to the desired output voltage, consistent with the constraint that, with a buck
converter, the end-of-life battery voltage must be above the required output voltage.
(For a boost converter, the constraint would be that the maximum battery voltage must
channel devices. On the basis of FET losses alone, it is desirable to choose a conversion
4.3 System-Level Considerations 120
ratio which ensures that current is carried by the NMOS device for a large fraction of
the cycle. For example, consider the CMOS buck topology drawn in Figure 3.2. For a
given output voltage and current, the losses in the power transistors are minimized if
the NMOS device carries the inductor current for the majority of the cycle. This calls
for a large conversion ratio, as far from 1:1 as possible. With a 5:1 conversion ratio, for
example, the PMOS device will conduct for only 20% of the cycle, and its losses can be
made small.
Thus, for conversion ratios near 1:1, it may be desirable to reconfigure the
buck topology as shown in Figure 4.27. In this circuit, the NMOS device functions as
the pass device, and, for conversion ratios near 1:1, it will have the longer conduction
interval. Similar reconfigurations of the boost and buck-boost topologies are possible to
minimize losses at extreme duty cycles. Figure 4.28 plots filter inductance and FET
losses versus conversion ratio for a buck circuit with fixed output voltage.
battery voltage should be selected as close as possible to the voltage at which the most
power is required, minimizing the size and maximizing the efficiency of the converter
supplying that voltage. The remaining converter topologies would then be chosen to
+ + +
Mp vx iLf Cf Vo
- -
Vin Cin
Lf Output
Mn Filter
-
Chopped Signal
L ∝ (1-D)
0
0 0.2 0.4 0.6 0.8 1
Conversion Ratio D
Fig. 4.28: Value of L and FET losses vs. conversion ratio for fixed Vo.
Linear regulators have the advantage that they are physically small and simple
result, one or more linear regulators are often integrated together (sometimes with their
for a cellular phone, which consists of five separate linear regulator outputs, is
possible, and would provide size and cost advantages similar to those enjoyed by
converter. In Section 4.2.5, it was shown that as the power transistor gate-drive supply
roughly as V g . There, it was also shown that a gate-drive voltage near 1.5 V is nearly
optimal in terms of total power transistor losses. (This is an encouraging result as many
modern-day digital ICs are trending toward such a voltage supply.) Thus, a low sub-
system voltage may be utilized as the gate-drive supply for each DC-DC converter in
amplifiers, and comparators − are often required to implement the PWM and/or PFM
control functions, power dissipation due to digital logic and control is becoming
increasingly important to the overall control power budget. It is these circuits which
can benefit most from supply voltage scaling (see Chapter 2).
microprocessor is reviewed. Here, separate sub-system voltages are used within each
converter. The lithium ion battery powers the analog blocks directly; a low-voltage
digital supply is used to reduce the dissipation of the digital control circuits; and a sub-
volt supply provides power for a high-speed low-swing I/O bus on the converter.
scaling (see Section 4.1.1), its losses increase. In a complete power delivery system
consisting of a number of DC-DC converters, frequency scaling may be used such that
the size and efficiency of different converters are traded, yielding the desired
4.3 System-Level Considerations 123
combination of overall size and losses. For example, the power supply with the highest-
power requirement may be optimized for high efficiency and reasonable size (with an
operating frequency in the hundreds of kHz), and all supplies with lower-power
requirements may be optimized for small size and reasonable efficiency (with operating
among converters which are integrated on the same die. Oscillators, reference voltages,
and master bias generators are needed in the control loop of any PWM or PFM DC-DC
particularly for a lower-power converter. If these components are shared among several
converters, the overall quiescent operating power and component count of the power
system will be reduced. In the design example of Chapter 6, successful resource sharing
Chapter 5
chapter, design considerations for the key DVS enabler, called the dynamic DC-DC
Section 5.1 introduces the principle of operation and key performance metrics
for a dynamic DC-DC converter. An example DVS system is shown in Section 5.2 and
followed throughout the chapter. In Section 5.3, the performance metrics are detailed.
Section 5.4 and Section 5.5 illustrate their impact on dynamic DC-DC converter and
overall DVS system design. Section 5.6 summarizes previous work on dynamic DC-DC
L fixed Vo
+ DC-DC +
1.0
3.3V
0.5
1.05V
0
0 0.5 1.0
Normalized load performance
Chapter 4. The load and static DC-DC converter do not communicate. The converter
A dynamic voltage scaling (DVS) system is shown in Figure 5.2. Here, the
performance and energy dissipation of the load are traded dynamically by varying the
5.1 Dynamic Converter Definitions 126
Request
L variable Vo
Dynamic
+ DC-DC +
Vbat C Load Vo
- -
Normalized load energy dissipation
1.0
3.3V
0.5
gV
o
in
as
cre
de
1.05V
0
0 0.5 1.0
Normalized load performance
Fig. 5.2: Dynamic voltage scaling for energy-efficient variable throughput processing.
converter output voltage at run-time. The dynamic DC-DC converter and its load must
The dynamic DC-DC converter has several requirements which differ from
those of the static DC-DC converter. While both converters must maintain a
substantially DC output voltage with high efficiency during regulation, the dynamic
DC-DC converter must do so over a much wider range of voltages and currents. In
addition, the dynamic converter must slew its output voltage during transitions at rates
approaching volts per microsecond, and must transfer large quantities of energy from
Vo or performance
regulates to ideal
DC output voltage
with 100% efficiency
majority of time in
lowest energy state
time
converter output voltage or load performance is plotted versus time. The DVS system
spends the majority of its time at the lowest voltage where it enjoys its largest energy
savings. Each voltage adaptation requested by the load is instant and lossless, providing
environmental conditions, load current, and battery discharge, and does so with 100%
power efficiency.
Figure 5.4 shows a practical DVS transient waveform. Voltage adaptations are
voltages are no longer maintained. The output voltage ripple causes extra load energy
dissipation. The converter itself dissipates regulation energy, and is likely to be least
efficient at the lowest output voltage and power levels − where efficiency matters most.
5.2 DVS System Example 128
non-zero
Tracking Time and Tracking Energy
Vo or performance
non-zero Voltage Ripple
causes load dissipation
time
Figure 5.5 shows a voltage and frequency tracking loop for use in a dynamic
voltage scaling system. The desired processor throughput is commanded by the process
frequency, f REF. The dynamic voltage converter consists of a frequency detector, a loop
filter, and the buck DC-DC converter described throughout this thesis and shown in
Figure 3.2. The frequency detector generates a digital error signal in proportion to the
frequency error, M ⋅ f REF – f VCO . This error is translated into an update signal for the
DC-DC converter through the loop filter. The DC-DC converter provides the voltage
supply, V dd , to the processor, regulating against changes in battery voltage and load
processor, and designed to match its critical path. The loop forces the output frequency
of the VCO, f VCO , to equal the commanded frequency M ⋅ f REF , at an input voltage V dd .
5.2 DVS System Example 129
Battery
DVS converter
Frequency
Detector L Vo fVCO
+ Loop Buck
Σ Filter Converter
VCO
1.05 V to > 3.3 V
− C
Vdd
Requested fVCO from process scheduler
µP <
The processor is therefore run at the minimum voltage supply, V dd , at which the
throughput request can be met, resulting in the lowest achievable energy per operation
The DVS system of Figure 5.5 has been prototyped and is further described in
Chapter 6. Some key system parameters, summarized in Table 5.1, are used for
5 MIPS at 1.05 V
There are two primary objectives of the dynamic DC-DC converter whose
• Minimize energy consumption of the entire DVS system for a given set of
demand
and the energy penalty associated with output voltage ripple, minimize tracking energy,
and minimize tracking time. In the following subsections each of these performance
metrics are detailed, and their impact on the buck converter design and the DVS system
To effect the large and rapid DVS transitions of its output voltage, the dynamic
DC-DC converter must efficiently transfer large quantities of energy from input to
output and vice-versa (Figure 5.6). The energy dissipated during these tracking
If the DVS voltage excursions are made through a resistive element alone,
E bat = ∆E C + Etrack
v o2
Vbat Dynamic Vo
v o1
Converter
C Load
bypass
1 2 2
∆EC = --- ⋅ C ⋅ ( v o2 – v o1 )
2
excursion from v o1 to v o2 from a battery at potential V bat (Figure 5.7). The change in
1 2 2
∆E C = --- ⋅ C ⋅ ( v o2 – v o1 ) (Eq 5-1)
2
∆Q C = C ⋅ ( v o2 – v o1 ) (Eq 5-2)
All of the charge delivered to the bypass capacitor is supplied by the battery. The
energy drawn from the battery is equal to the product of the delivered charge and the
Linear regulator
Vbat Vo
1
E track = ∆E bat – ∆E C = --- ⋅ C ⋅ ( v o2 ⋅ ( 2V bat – v o2 ) – v o1 ⋅ ( 2V bat – v o1 ) ) (Eq 5-4)
2
With V bat = 3.6 V, v o1 = 1.05 V, and v o2 = 3.3 V, E track = 3.21 µJ per 1 µF of bypass
capacitance. For the DVS system parameters of Table 5.1, this is equal to the energy
If the voltage excursions are instead made through an ideal DC-DC converter,
they are ideally lossless. This is because the large input to output voltage ratio is
applied across a series inductor, rather than a resistor. Figure 5.8 shows equivalent
circuits and several cycles of i L (t) and v o (t) waveforms for a tracking voltage transition
(V bat -Vo ) and -Vo , storing and releasing energy from the battery to the output capacitor.
the tracking transition. In the integrated DC-DC converters described in Chapter 4, gate
and switching losses, resistive losses, and control power can usually be kept to below
10% of the energy handled by the converter. Using this conservatively high relative
dissipation together with (Eq 5-1), a simple first-order estimate to the tracking energy
of a DC-DC converter may be found. For example, with v o1 = 1.05 V and v o2 = 3.3 V,
E track = 0.49 µJ per 1 µF of bypass capacitance − only 15% of the energy required by
the linear regulator. This is equivalent to the energy consumed by 1225 instructions at 5
iL iL
+ L + L +
Vbat C vo C vo
− − −
iL(t)
vo(t)
Fig. 5.8: A DVS tracking transition through a discontinuous mode DC-DC converter.
transition into a low-power mode. For every 1 µF of output bypass, 0.49 µJ is dissipated
only worthwhile if the energy saved by computing at low-voltage is greater than the
0.98 µJ
---------------------------------------------------------- = 239 instructions at 5 MIPS (Eq 5-5)
4.5 nJ/inst. – 0.4 nJ/inst.
is the break-even point. For fewer than 239 low-voltage instructions, it is more energy-
according to (Eq 5-1) the value of output bypass capacitor, C, should be minimized.
5.3 Dynamic DC-DC Converter Performance Objectives 134
The tracking time of the dynamic DC-DC converter determines the latency
throughput mode. A fast transition from the most energy-efficient mode (the lowest
output voltage) to the highest throughput mode (the highest output voltage) is critical to
applications.
Figure 5.9 shows the fastest possible transition of the converter output voltage. A large-
signal step-response of state variables i L and v o to a full 3.6 V step at the inverter
output node, v x , is plotted for L = 3.5 µH, C = 5 µF, and R = 0.12 Ω. The time constant
Series LCR step respsone. L = 3.5 uH, C = 5 uF, R = 0.12 Ohms, Vin = 3.6 V
3
1
iL [A]
−1
−2
−3
0 0.2 0.4 0.6 0.8 1 1.2
−4
x 10
4
vo [V]
1
0 0.2 0.4 0.6 0.8 1 1.2
time [sec] −4
x 10
Thus, to minimize tracking time, small-valued output filter elements are desired.
current of:
v o2 – v o1
ˆi = ----------------------
Lf (Eq 5-7)
L⁄C
dynamic DC-DC converter is usually necessary to avoid damage to the converter IC and
filter elements during tracking transitions. When the inductor current is limited to some
I max , the maximum slew rate during output voltage transitions is:
dV I max
---------o- = ----------
- (Eq 5-8)
dt C
environmental conditions. During this regulation mode, the dynamic converter behaves
similarly to the more conventional static DC-DC converter, and can operate with
Figure 5.10 shows the general trend of decreasing DC-DC converter efficiency
at lower output voltage and power levels − exactly where the converter must be
5.3 Dynamic DC-DC Converter Performance Objectives 136
pwm
pfm
20
10
0
1 1.5 2 2.5 3 3.5
Vo [Volts]
efficient to most effectively conserve energy in a DVS system. Here, P out scales with
Vo 3 , similarly to a DVS system, V bat = 3.6 V, output inductor and capacitor values (L =
3.5 µH, C = 5 µF) are chosen from the example DVS system of Section 5.2, and power
transistor sizes are optimized for the 3.3 V, 450 mW operating point. Quiescent control
currents of 500 µA in PWM mode and 100 µA in PFM mode are assumed. Total power
FET dissipation, conduction loss in the external filter elements, and static control
power are considered for a continuous conduction mode PWM converter and a constant
peak current controlled PFM converter. From this plot, it is clear that some type of PFM
lower-voltage, lower-current operating points. (This was also the conclusion of Section
4.2.7.)
Figure 5.11 and Figure 5.12 show the contribution of three key mechanisms of
loss relative to P out in PFM and PWM operation. Three important sets of observations
Quiescent control
Power FET
Inductor Conduction
2
0.5
0
1 2
Quiescent control
Power FET
Inductor Conduction
10
Normalized losses as percentage of Pout
0
1 2
• In both PWM and PFM modes, total power FET dissipation increases substantially
relative to load power with decreasing output voltage. Dynamic transistor sizing
as a function of Vout (described in Section 4.2.4) will help to mitigate this effect.
• In both PWM and PFM modes, the relative contribution of static dissipation from
trend is more noticeable under PWM operation since the control circuits are
generally more complex and power hungry than their PFM counterparts. A PFM
controller, whose power dissipation scales with the load, will generally
to scale the power dissipation of the controller together with the load.
• In PWM mode, conduction loss through the series resistance of the output filter
mode, conduction and switching losses usually scale relative to the output power
3
because the PFM repetition period scales with load. However, since P out ∝ V o ,
In Chapter 3, it was shown that the output of any DC-DC converter includes a
symmetric AC ripple voltage superimposed on the desired DC. The magnitude of this
ripple was given in (Eq 3-3) for PWM operation and in (Eq 3-26) for PFM operation
V o ⋅ ( 1 – V o ⁄ V bat )
∆V PWM = ----------------------------------------------- (Eq 5-9)
P→P 2
8 ⋅ L ⋅ C ⋅ fs
5.3 Dynamic DC-DC Converter Performance Objectives 139
2
1 T pmos ⋅ ( V bat – V o ) ⋅ V bat
∆V PFM = --- ⋅ ---------------------------------------------------------------- (Eq 5-10)
P→P 2 Vo ⋅ L ⋅ C
ripple scales inversely with the square of the LC time constant. (Eq 5-9) and (Eq 5-10)
also indicate that normalized output voltage ripple, ∆V / Vo , increases for decreasing
Vo . Figure 5.13 shows the trend. Here, inductor and capacitor values are fixed, V bat =
3.6 V, and normalized output voltage ripple is plotted versus DC output voltage, Vo .
general purpose processor of the DVS system, particularly at low voltage. A simple
first-order model is used here to estimate the impact. In Figure 5.14, the peak-to-peak
output voltage ripple, ∆V, is symmetric about the desired DC output voltage, V nom . The
required processor throughput, f nom , is maintained by dithering between high and low
pfm
pwm
7
6
Normalized output voltage ripple
0
1 1.5 2 2.5 3 3.5
Vo [Volts]
duty cycle, d
fhi
Throughput fnom
dithers such that fave = fnom
flo
Fig. 5.14: First-order frequency and voltage model used to compute voltage ripple energy.
V nom +∆V/2 to V nom -∆V/2. The resulting average system throughput is f nom . The
Figure 5.15 plots the normalized load energy dissipation, E ave ⁄ E V , for
nom
various normalized output voltage ripple, ∆V ⁄ ( 2V nom ) . This data assumes that energy
and delay scale with voltage as shown in Figure 2.1. From this plot, a zero-to-peak
output voltage ripple as high as 5% might be considered tolerable, even at low nominal
voltages. Higher normalized output voltage ripple may be acceptable above 1.5 V. For
According to (Eq 5-9) and (Eq 5-10), this indicates the use of large values of L and C,
1.45
1.40
1.35
Normalized energy
1.30
1.25
1.15
1.10
10% zero-peak Vo ripple
1.05 5% zero-peak Vo ripple
1.00
1.0 1.5 2.0 2.5 3.0 3.5 4.0
Vo(nom) [V]
Fig. 5.15: The impact of output voltage ripple on load energy dissipation.
There are two sets of performance metrics that trade-off in a dynamic DC-DC
converter through the values of output filter inductor and capacitor. With larger values
of L and C, the converter is a better regulator, with improved conversion efficiency and
reduced output voltage ripple. With smaller values of L and C, the converter is a better
objective. If most operations are performed at a low throughput and low energy corner,
tracking transitions are rarely made, and latency can be tolerated, large values of L and
5.5 Impact of Performance Metrics on System Performance 142
converter losses.
The need for high conversion efficiency over a wide range of output voltage
3
and power levels, with P out ∝ V o , presents a difficult design challenge. A hybrid PWM-
PFM control scheme is most likely to maintain low-voltage efficiency by ensuring that
the converter switches only when necessary. Dynamic transistor sizing is necessary to
minimize power FET losses over the output range of the converter. The voltage scaling
techniques (see Chapter 6), must be judiciously used to scale the quiescent control
DC-DC converter. Here, the impact of these nonidealities on DVS energy savings is
quantified.
Assuming a 95% DC-DC converter efficiency at 3.3 V and 450 mW, the
converter-load combination of Figure 5.5 dissipates 4.7 nJ per instruction at the 100
MIPS at 3.3 V operating point 1 . With 5 µF of output bypass capacitance and 10%
tracking energy dissipation, the dynamic converter dissipates a total of 4.9 µJ for one
complete 3.3 V to 1.05 V to 3.3 V tracking cycle. At the low energy operating point − 5
MIPS at 1.05 V − the processor dissipates 0.4 nJ per instruction. For a DC-DC
converter with a 5% zero-to-peak output voltage ripple at 1.05 V, Figure 5.15 indicates
a 4% energy dissipation penalty in the load. With an 85% conversion efficiency at 1.05
1. Extra load energy dissipation due to output voltage ripple is negligible at 3.3 V.
5.5 Impact of Performance Metrics on System Performance 143
0.4 nJ ⋅ ( 1 + 0.04 )
E diss = ------------------------------------------- = 0.5 nJ/instr (Eq 5-13)
0.85
Figure 5.16 plots the normalized energy dissipation (E DVS /E 3.3V ) as a function
4.9 µJ
N 1.05V = ----------------------------------------------------------- = 1167 instructions (Eq 5-14)
4.7 nJ/instr – 0.5 nJ/instr
run-time at 5 MIPS.
1.2
1.1
0.9
EDVS / E3.3V
0.8
0.7
0.6
0.5
0.4
0.3
1000 1500 2000 2500 3000 3500 4000 4500 5000
Number of instructions
Each of the approaches is similar to the block diagram of Figure 5.5, in that the critical
path of the processing element is replicated by a delay element or ring oscillator, and
designs for use in dynamic voltage scaling systems. Each identifies conversion
efficiency over a wide range of output voltage and power levels to be of primary
output, are implemented in an effort to scale the quiescent dissipation of the converter
with decreasing output voltage. Table 5.2 benchmarks the three designs.
L 50 µHa 7 µH 8 µH
C 50 µFa 33 µF 32 µF
[Kuroda98] uses a continuous conduction mode PWM control scheme for the
conversion efficiency drops to 40% at the lowest voltage and power levels (10 mW at
0.8 V). L = 8 µH and C = 32 µF appear to be chosen for low output voltage ripple,
quoted at less than 0.1% at a constant load current. A first-order tracking response is
observed with a time constant on the order of 40 µs, which is set by an integral term in
the controller. Tracking energy is not documented, but can be estimated to exceed 32 µJ
efficiencies between 82% and 92% over a 2.0 V to 4.5 V output voltage range when
driving a constant resistive load of 25 Ω. Inductor and capacitor values are not
provided, but a published bode plot shows an LC corner frequency below 3 kHz. For
large values of output filter elements, low output voltage ripple is guaranteed. Tracking
time is dominated by a pair of pole-zero doublets from the LC output filter and PID
value of inductance (L = 7 µH). The converter output is regulated using a constant on-
time PFM control scheme, and therefore requires a fairly large output capacitor (C = 33
µF) to guarantee low ripple at the lowest output voltages. High conversion efficiency is
maintained over a wide range of output voltage and power levels: The maximum
efficiency is 93% at the 3.5 V, 200 mW operating point; the minimum efficiency is 83%
2. Assumes 20% energy loss in the converter; loosely estimated from published efficiencies.
3. Assumes 10% energy loss in the converter; loosely estimated from published efficiencies.
5.6 Summary of Previous Work 146
to 3.3 V to 1.05 V output excursion 4 . Tracking time is listed at better than 6 ms/V, but
second-order response.
differentiated from previous work on the basis of its increased dynamic range (1.05 V
to greater than 3.3 V output voltage and less than 1 mW to 500 mW output power),
dynamic voltage scaling systems, particularly as their performance requests vary with
greater frequency.
4. Assumes 10% energy loss in the converter; loosely estimated from published efficiencies.
147
Chapter 6
Prototype DC-DC
Converters
at both the system and circuit levels. Power system design decisions are documented in
Section 6.1. All high-speed communication from the processor to its peripherals is
made via the low-swing I/O circuits described in Chapter 2, yielding as much as a 275x
reduction in power. A small DC-DC converter provides a regulated voltage supply (with
an adjustable output from 0.1 V to 0.5 V) to power the I/O transmitters. Section 6.2
details the design of this ultra-low-voltage DC-DC converter. The processor core and
achieving throughputs ranging from 100 MIPS at 3.3 V, to 5 MIPS at 1.05 V. At the 5
MIPS operating point, a nearly 10x improvement in energy per operation is achieved.
The dynamic DC-DC converter which enables this DVS scheme is described in Section
6.3.
converter is designed to provide a 1.5 V output at a 500 mA full load current. It was
6.1 Processor Power Delivery System 148
communication from the processor to its memory and peripherals is accomplished via
measured results are documented in Section 6.2, enables the transceivers, delivering a
regulated supply (adjustable from 0.1 V to 0.5 V) with an efficiency as high as 85%.
energy consumption. The design and measured results of the dynamic DC-DC converter
The remainder of this section contains a discussion of the power system design
The DC-DC converters leverage the three available power supply voltages for
improved energy efficiency. While all analog circuits must operate from the battery
voltage for headroom considerations, critical digital hardware is operated from the
dynamically scaled voltage supply. This scales the energy per operation of the
converters’ digital control circuits together with those of the microprocessor and
6.1 Processor Power Delivery System 149
L1
0.2 V
DC-DC
Lithium Ion C1
Cell
+
3.6 V
nominal
L2 1.05 V to > 3.3 V
-
DVS
C2
Battery:
Vbat(max) = 4.2 V
Vbat(nom) = 3.6 V
Vbat(min) = 3.0 V
Processor system:
Request
100 MIPS, 4.5 nJ/inst. at 3.3 V
5 MIPS, 0.4 nJ/inst. at 1.05 V Memory
Full-speed 32-bit bus (50 pF per bit) Processor and
32 bits Peripherals
DC-DC Converter:
Adjustable 0.1 V to 0.5 V output
> 80% efficiency Low-swing bus drivers
L1 = 10 µH, C1 = 20 µF
DVS Converter:
Dynamic 1.05 V to > 3.3 V output
20 µs tracking time
> 85% energy efficiency
L2 = 3.5 µH, C2 = 4.7 µF
utilized by the dynamic voltage converter to enable low-swing communication from the
static power. Since the power of one master bias can be amortized over two converters,
the light-load efficiencies of each are improved. In addition, the 4 MHz DVS system
subsystem itself; thus, its power consumption does not count against either converter.
The initial plans for implementation of this power system included the highest
levels of integration. However, integration of the converters together with the processor
load is deemed infeasible due to the large voltage transients on the power FET ground
lines. Since an epi process has been chosen for fabrication of the processor, sufficient
isolation of the power FET ground noise from the processor circuits cannot be
guaranteed.
technically feasible since all of the high-current power FET switching transitions are
synchronized to the same system clock. However, for testability, the two chips were
The processor subsystem of Figure 6.1 includes a full-speed 32-bit bus. If this
bus were switched at the full DVS output voltage, it would approximately double the
system energy per instruction. Assuming a 25% activity factor and 50 pF per bit, E bus =
4.4 nJ/inst at 100 MIPS and 3.3 V, and 0.4 nJ/instr at 5 MIPS and 1.05 V.
6.2 An Ultra-Low-Voltage DC-DC Converter 151
The analysis presented in Section 2.4 shows the energy savings effected by
powering the bus transmitters from the battery through an ultra-low-voltage DC-DC
V dd
β = ----------- (Eq 6-2)
V LO
1), even for very low efficiencies, the converter-load system results in a more energy-
efficient bus transmitter. For η = 0.7 and V LO = 0.2 V, E bus is reduced to 20 pJ/instr −
an addition of only 5% to the processor energy per instruction at 5 MIPS and 1.05 V. At
100 MIPS and 3.3 V, E bus adds less than one half of one percent to the processor energy
per instruction.
The primary challenges to the DC-DC converter design are the ultra-low
output voltage and current levels that must be supported with reasonable efficiency.
Therefore, all control system, architecture, and circuit-level decisions are made with
constant on-time, synchronous PFM control scheme. By exploiting the existing 4 MHz
DVS system clock, this simple controller offers ultra-low static power dissipation. A
block diagram of the controller is shown in Figure 6.2. The system timing diagram is
Vref - VLO
comparator Vbat
−
driver
Vref + S
Q iL
250 ns R
VLO = 0.2 V
sampler L
driver
S
edge detector Q C LOAD
R
− +
NMOS current
comparator
The system is synchronized to the existing 4 MHz DVS reference clock. At the
0 ns edge, the Vref −V LO comparator bias is enabled. 125 ns later, the comparator
sampling and pre-amplification switches are sequenced to initiate the comparison of the
converter output, V LO , to a low-power external reference, Vref . At the 250 ns edge, the
comparator output is sampled. If Vref < V LO , the power PMOS device is left off, the
If Vref > V LO , the PMOS device is turned on (always on the 250 ns edge) and
conducts for T pmos = 250 ns. During this interval, inductor current i L ramps linearly
from zero to its peak value, I p . When the PMOS turns off (always on the 500 ns edge),
feedback timing control turns the NMOS device on to pick up the inductor current.
conduction time interval may be found relative to the PMOS conduction interval. This
( V bat – V LO )
T nmos = T pmos ⋅ --------------------------------- (Eq 6-3)
V LO
6.2 An Ultra-Low-Voltage DC-DC Converter 153
Vo-cmp bias
vgp
vgn
iNMOS-cmp 4 clocks
bias
iNMOS-cmp
output enable
Noff Tnmos
iL 0
Vref
VLO
Vo < Vref
deliver charge Vo > Vref Vo > Vref
idle for 4 us idle for 4 us
The minimum interval, T nmos = 1.25 µs, is found at the 3.0 V minimum battery voltage
and the 0.5 V maximum output voltage. To conserve energy, the NMOS current
comparator bias is not enabled until 1 µs after the NMOS device is gated − the 1.5 µs
edge. The maximum interval, T nmos = 10.25 µs, is found at maximum battery (4.2 V)
and minimum output (0.1 V), and sets the upper limit to NMOS current comparator
energy dissipation. The comparator bias is given 250 ns to settle; its output is not
monitored until the 1.75 µs clock edge. The NMOS is turned off asynchronously by the
NMOS current comparator when i L has decayed to zero. The cycle then repeats, re-
The control system has been verified using matlab simulation. Figure 6.4 and
Figure 6.5 show the start-up transient and steady-state operating waveforms for L = 10
vx [V]
2
0
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
100
50
iL [mA]
−50
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
300
vo [mV]
200
100
0
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
time [ms]
4
vx [V]
0
0.6 0.65 0.7 0.75 0.8
100
50
iL [mA]
−50
0.6 0.65 0.7 0.75 0.8
210
vo [mV]
205
200
195
0.6 0.65 0.7 0.75 0.8
time [ms]
For lowest static power dissipation, the DC-DC converter exploits the existing
4 MHz clock and 10 µA master bias of the DVS system. While for headroom
considerations, analog components must run from the full lithium ion battery source
voltage, V bat , all critical digital hardware is supplied by the dynamically scaled voltage,
The digital master control synchronizes the system to the 4 MHz clock, and
negotiates power-on reset sequencing. From Figure 6.3, it provides clocks for the Vref −
V LO comparator, triggers the i NMOS comparator sequence, and commands the power
transistors. At the core of the master control are 16 low-voltage TSPC registers and
various combinational logic which provide clock division and generation functions.
Figure 6.6 shows the positive edge triggered TSPC register with level-sensitive reset
input. Device sizes have been modified from the UC Berkeley low-power library to
allow operation down to 1.0 V at the 3σ fast NMOS / slow PMOS process corner.
CLK RST
6/2
4/2 19/2 p: 15/2
n: 5/2
Q
6/2
D CLK CLK
4/2 4/2
4/2
p: 3/3
n: 3/11
CLK
4/2 4/2 3/2 4/2
Fig. 6.6: TSPC register with operation to 1.0 V. Device sizes in λ = 0.3 µm.
6.2 An Ultra-Low-Voltage DC-DC Converter 156
Figure 6.7 shows the level converter circuit from the DVS output voltage, V dd ,
to the battery voltage, V bat , required by the comparator clocks and the power
transistors. Device sizes have been chosen to allow operation down to 1.0 V at the 3σ
During idle cycles (Vref > V LO ) the 4 MHz system clock is divided into a 16-
phase period at V dd . Once per idle cycle, the Vref −V LO comparator clocks are sequenced
by the master control at V bat . The effective switched capacitance during an idle cycle
is:
• C Vdd = 8.8 pF at V dd
For V dd = 1.05 V and V bat = 3.6 V, this contributes 4.0 µW of static power dissipation.
During active cycles (Vref < V LO ) the 4 MHz system clock is divided into a 7-
phase period at V dd . The master control provides the Vref −V LO comparator clocks, the
power transistor control signals, and triggers the i NMOS comparator. The effective
• C Vdd = 3.1 pF at V dd
For V dd = 1.05 V and V bat = 3.6 V, this contributes 21.5 pJ of energy dissipation.
Vbat
Vdd
3/2 3/2 in 0
outb out
Vbat
in 38/2 38/2 inb
out 0
comparator is similar to those described in [Yin92] and [Lynn95] and is known to have
a good combination of speed and accuracy with relatively low power dissipation.
The comparator switch sequence is shown in Figure 6.9. At rest, ΦBIAS is low,
ΦEQ is high, and ΦEVAL is low. The bias current is disconnected from the
aVdd swVdd
15/2 15/2
VLO 300/5 300/5 Vref
ΦEQ
S Q 5x U_Vref
Y X
4/5
15/2 15/2 R
Qb
aGND aGND
Fig. 6.8: Vref−VLO comparator schematic. Device sizes given in λ = 0.3 µm.
clk4
ΦBIAS
ΦEQ
ΦEVAL
tcmp
U_Vref
0 ns 125 ns 250 ns
preamplifier, nodes X and Y are shorted through a switch, and the cross-coupled PMOS
load is disabled. The comparator sequence is initiated by the master control at the “0
ns” edge when ΦBIAS is asserted, enabling the preamp. The equivalent preamplifier
circuit is shown in Figure 6.10. Here, the shorting switch is sized to ensure that the M3-
M4-R O positive feedback loop gain is less than one over process, temperature, and
V XY gm 1 R O
A V = ---------------------------- = --------------------------- (Eq 6-4)
V ref – V LO 2 – gm 3 R O
At the 125 ns edge, the master control lowers ΦEQ, and after a short non-
overlap interval, asserts ΦEVAL. The shorting switch is released, and V XY is amplified
and latched by the cross-coupled PMOS and NMOS loads to full digital levels. The
nand-based SR flip-flop generates a digital signal, “U_Vref”, which is high when the
Vdd
30 µA
VLO Vref
M1 M2
RO
Y X
M3 M4
positive feedback
(loop gain < 1)
converter output voltage falls below Vref . This output is sampled by the master control
Circuit simulation results are shown in Figure 6.11. Static current consumption
evaluation. Here, an unanticipated short circuit path increases the static current
LVcmp.tr0
LVcmp.tr0
phibias
phieval
I(vbat)
u_vref
vref
phieq
vo
5.00u
5.00u
4.00u
4.00u
TIME (S)
TIME (S)
3.00u
3.00u
2.00u
2.00u
XP 1997.202, (c) 1997 Avant! Corporation
1.00u
1.00u
0.
-16.15m
-2.15m
1.71m
0.22
0.21
0.21
0.20
0.20
0.19
3.60
3.60
3.60
3.60
0.
0.
0.
3.6 V 3.6 V
3.6 V
30 uA
0.20 V 3.48 V
1.78 V
0.19 V 0.20 V
1.95 V 0.20 V
consumption by a factor of nine. The short circuit path is present only during
evaluation. It follows the direction of decreasing bias voltages as shown in Figure 6.12.
Unfortunately, this problem was not discovered until after tape-out, and caused a
differential input signal, the comparator is highly overdesigned. The total energy
rectifier when i NMOS = 0 from above (Figure 6.3). Chapter 3 discusses the energy
6.2 An Ultra-Low-Voltage DC-DC Converter 161
0.025
0.02
0.015
0.01
0.005
0
0 2 4 6 8 10 12 14 16 18 20
abs(Ierror) [mA]
dissipation penalty associated with early and late NMOS turn-off (see (Eq 3-42) and
(Eq 3-43)). Figure 6.13 illustrates the results: Energy dissipation penalty normalized to
the energy delivered to the load in a single PFM burst is plotted as a function of the
comparator and gives some design guidance. For 1 mA of error in either direction, the
2
energy penalty is 5.0 pJ. This scales as ( I ε ⁄ 1 mA ) until I ε is large enough to induce
body diode conduction. Since the nominal energy delivered to the load during a single
PFM pulse is 38.2 nJ, this extra dissipation starts becoming important for I ε on the
order of 10 mA and above. Note from the plot that for such values of I ε , body diode
conduction is indeed induced. Here it is desirable to gate the NMOS device a little late,
rather than early. This is due to the fact that at such a low output to battery voltage
6.2 An Ultra-Low-Voltage DC-DC Converter 162
ratio, the PMOS body diode dissipates less energy than the NMOS body diode for equal
conduction intervals.
i NMOS is inferred through the voltage drop across the NMOS channel:
Since R N is as small as 75 mΩ, low offset voltage is a primary design consideration for
the comparator. For Vos = 1 mV, the equivalent i NMOS error is:
1 mV
I ε = ----------------- = 13.3 mA (Eq 6-6)
75 mΩ
di NMOS – V LO – 0.2 V
- = ---------------- = – 20 µA ⁄ ns
-------------------- = ------------- (Eq 6-7)
dt L 10 µH
comparator is designed to consume energy only during PFM pulses. A strobed bias
[Acker95]. Two input-offset cancelled differential amplifier stages form the main
preamplifier. A high gain differential to single-ended amplifier and a nand gate convert
the output to full-swing digital levels. In reset mode, the bias to the amplifiers is
disabled, switches phased Φ1, Φ2, and Φ3 are closed, and switches phased Φ4 are open.
To conserve static power, the master control does not enable the comparator bias until
exactly 1 µs after the power NMOS device is gated. In the succeeding 250 ns, the
6.2 An Ultra-Low-Voltage DC-DC Converter 163
Φ1 Φ2
Φ3 diffal to se
Φ4 C1 C3
+ + - + - +
Vin Φ4 Noff
- - + - + -
C2 C4
Φ3 Φout
Φ1 Φ2
digital output enable
1st gain stage 2nd gain stage
preamplifier offset is stored on the interstage coupling capacitors. (The input capacitors
also serve to level-shift the inputs, extending the input common-mode range below
ground.)
Compare mode is entered in the sequence shown in Figure 6.15. The release of
reset mode is initiated by the master control after 250 ns of offset storage by opening
switches phased Φ1. Any charge injection mismatch into C1-C2 due to the opening of
switches phased Φ1 is amplified by the first stage and stored differentially on C3-C4.
Switches phased Φ2 and Φ3 are then released, and Φ4 is closed, connecting the power
NMOS drain and source terminals to the comparator input. Differential voltage stored
on the capacitors now subtracts from the input voltage, cancelling the offset voltages of
the preamp stages and any charge injection mismatch. The effective input-referred
∆Q
V os,eff = V OS2 + ---------- ⋅ --------------------------------
1
(Eq 6-8)
C 3,4 A 1 ⋅ ( 1 + A 2 )
When v dsNMOS crosses zero, the i NMOS comparator gates the power device and disables
6.16. Device sizes have been chosen to maximize gain-bandwidth product. Tail current
6.2 An Ultra-Low-Voltage DC-DC Converter 164
iLF 0
vgn
Noff_bias_S
Noff_Φ1_R
ΦBIAS
Φ1
Φ2
Φ3
Φ4
Φout
Noff
C1-C2 ensure that kT/C noise has a negligible impact on comparator accuracy. 200 pF
injection mismatch error and capacitive loading on the first differential gain stage.
6.2 An Ultra-Low-Voltage DC-DC Converter 165
aVdd
gmn ( gmp + gon )
gain = ----------- BW = --------------------------------
Mp Mp gmp CL
I1 I1
+
vo
−
Mn Mn
+ CL CL
vi AMP1 AMP2
−
Mn 24/0.6 21/0.6
IB Mp 1.2/0.9 1.5/1.2
IB 50 µΑ 30 µΑ
I1 10 µΑ 10 µΑ
aGND
Fig. 6.16: Differential gain stage in iNMOS comparator (device sizes in microns).
differential pair with a PMOS mirrored load. The stage has high gain ( g m ⋅ r o ) and is
The overall comparator has a linear gain of greater than 80 dB from input to
nand gate input. The simulated delay is less than 40 ns over process and battery
discharge, introducing an error of less than 0.8 mA into the system. With offset
cancellation, the 3σ input-referred offset voltage is estimated to be less than 0.5 mV,
introducing a worst-case error of 6.7 mA. The worst-case energy dissipation penalty
mirroring). The digital clock sequencers and switches yield an effective switched
capacitance of 3.5 pF per PFM pulse. The overall energy dissipation of the i NMOS
2
Ei = 80 pJ + ( 3.5 pF ) ⋅ V bat + V bat ⋅ ( 140 µA ) ⋅ ( T nmos – 1 µs ) (Eq 6-9)
NMOS
6.2 An Ultra-Low-Voltage DC-DC Converter 166
(Eq 6-9) includes the energy dissipation penalty introduced by a worst-case early or late
NMOS turn-off. For V bat = 3.6 V, T nmos = 4.25 µs, and E i = 1.76 nJ − 4.6% of the
NMOS
The converter exploits the 10 µA DVS system master bias for low-power. A
circuit schematic is shown in Figure 6.17. A digital signal strobes the 10 µA current
mirror. When “disable” is high, all current sources are cut off, eliminating static
dissipation. When “disable” is low, triode PMOS degeneration devices improve current
source output resistance. Simple cascode current sources are avoided for headroom
considerations. High-swing cascode current sources require extra mirrors and are
In idle cycles, the master mirror is enabled for 250 ns of the 4 µs period,
consuming an average current of 625 nA. During PFM pulses, the mirror is active
For V bat = 3.6 V and V LO = 0.2 V, T nmos = 4.25 µs and E bias = 171 pJ per PFM pulse.
Vbat
disable
10/20
40/10
biasIN
PAD
Vbat Vbat
C1 R1
Vref Vref
C2 R2
V bat V bat
V ref = ---------------- V ref = ----------------
C2 R1
1 + ------- 1 + -------
C1 R2
2
C1 ⋅ C2 2 V bat
P static = --------------------- ⋅ V bat ⋅ f P static = ---------------------
C1 + C2 R1 + R2
For Vbat = 3.6 V, Vref = 0.2 V: For Vbat = 3.6 V, Vref = 0.2 V:
C1 = 1 pF R1 = 1.7 MΩ
C2 = 17 pF R2 = 100 kΩ
P = 3.1 µW (for f = 250 kHz) P = 7.2 µW
divider may be used to generate the voltage reference, Vref , from the battery source
resistor divider based reference is implemented at the board level in the prototype
converter. This approach provides the additional advantage that the reference can be set
to a continuous range of values by using potentiometers for R1 and/or R2. For the
values given in Figure 6.18, the voltage reference consumes 7.2 µW of static power
+ Wp = 12.3 mm
Lp = 0.6 µm
L = 10 µH
Wn = 25.1 mm RC = 0.09 Ω
Ln = 0.6 µm Vo = 0.2 V
C = 20 µF
- -
component values, approximate parasitics, and device sizes, is shown in Figure 6.19.
Filter element values are selected according to the design equations of Chapter 3. L =
With V bat = 3.6 V, V LO = 0.2V, and T pmos = 250 ns, (Eq 3-22) gives:
2
1 ( 250 ns ) ⋅ ( 3.6 V – 0.2 V ) ⋅ ( 3.6 V )
Q L = --- ⋅ ---------------------------------------------------------------------------------------- = 191 nC (Eq 6-12)
2 ( 0.2 V ) ⋅ ( 10 µH )
mV of voltage sag.
Power transistors are sized according to (Eq 4-11) to minimize total energy
dissipation in the nominal 3.6 V to 0.2 V case. Tapering factors of 10-12 are used in the
dissipation per PFM pulse in a 3.6 V to 0.2 V application. The output inductor is the
delivered from battery to output per pulse, an overall power train efficiency of 86% is
achieved.
Total 6.2 nJ
Avanti’s Starsim simulator. Figure 6.20 shows the key circuit waveforms. Table 6.2
With V bat = 3.6 V, V LO = 0.2 V, and the processor operating at 100 MIPS and
V dd = 3.3 V, the average load at the regulator output is 8 mA. Here, the DC-DC
converter delivers power at a respectable 80% efficiency, with the power train
dominating the loss. The converter and low-swing I/O load combination consumes 2.0
mW from the 3.6 V battery. From (Eq 6-1) and (Eq 6-2), the converter enables a bus
energy dissipation reduction to only 0.4% of its original value at the high throughput
operating point.
6.2 An Ultra-Low-Voltage DC-DC Converter
XP 1997.202, (c) 1997 Avant! Corporation
*** test entire low-voltage regulation system : 98/06/29 00:21:12
1.50
LVreg.tr0
0. clk4
Fig. 6.20: Low-voltage regulator circuit simulation waveforms.
3.61
-69.97m noff
3.62
-15.59m noff_bias_s
3.62
-16.72m noff_phi1_r
3.68
-37.92m phieval
3.61
-19.42m u_vref
3.60
-15.32m vgn
3.60
-0.80m vgp
3.75
-48.13m vo_bias
0.21
1.68n vout
86.21m
-2.38m i1(lf)
2.00u 4.00u 6.00u 8.00u
1.00u 3.00u 5.00u 7.00u 9.00u
TIME (S)
170
printed at 18:40:04 Dec 3, 1998 by anthonys
6.2 An Ultra-Low-Voltage DC-DC Converter 171
delivers 80 µW at 56% efficiency. At this operating point, the converter enables overall
I/O dissipation reduction to 6.5% of its original value. Here, the power dissipation is
dominated by the Vref −V LO comparator. Had this comparator been designed for lower
power and more suitable delay, and if its short circuit path were eliminated, conversion
efficiency could be brought above 70% at the 5 MIPS and 1.05 V operating point.
The standby power of the converter is only 47.3 µW, and is also dominated by
the Vref −V LO comparator. This could be brought below 20 µW with a properly designed
comparator.
metal process in May, 1997. Figure 6.21 shows a chip plot, with 0.9 mm by 1.8 mm die
dimensions. The upper portion of the IC contains the PMOS and NMOS power
transistors, drivers, and 400 pF of tuned bypass capacitance. The bottom portion of the
6.2 An Ultra-Low-Voltage DC-DC Converter 172
chip includes the digital control and analog circuits. Separate supplies with local on-
chip bypassing are maintained for the power, digital, and analog components on the IC.
The chip is housed in a 16-pin DIP, with 2 pins each dedicated to VX, power
FET supply, and power FET ground. Double bonds are used to reduce the resistance of
Full functionality of the regulator was achieved over the full battery voltage
(3.0 V to 4.2 V), dynamically scaled digital supply voltage (1.05 V to 3.3 V), and
output voltage (0.1 V to 0.5 V) ranges. Due to the long lead-time on 10 µH inductors, L
iNMOS comparator
design objective. This also results in an increase in PFM pulse frequency, and a
corresponding increase in switching, gate-drive, and control losses relative to the load
consumption.
Figure 6.22 and Figure 6.23 show medium load and light load steady-state
operating waveforms for V bat = 3.0 V, V dd = 1.5 V, and Vout = 0.2 V. Peak-to-peak
output voltage ripple is kept below 4.5 mV in either case, which is consistent with the
v x and i L detail is shown at the NMOS zero-current turn-off transition. The NMOS
power FET turns off with i L = -7.6 mA, introducing 433 pJ of loss, or 2.1% (with L = 15
µH). This error is acceptable, and near the value predicted in Section 6.2.2.3.
3.0 V, V dd = 1.5 V, with Vout = 0.2 V and Vout = 0.5 V (Figure 6.25). In general, these
iL
20 mA/div
vo (AC)
10 mV/div
vx
2 V/div
Fig. 6.22: Medium load steady-state operating waveforms. Vbat = 3.0 V, Vdd = 1.5 V, Vout = 0.2 V,
Iload = 3.0 mA.
6.2 An Ultra-Low-Voltage DC-DC Converter 174
iL
20 mA/div
vo (AC)
10 mV/div
vx
2 V/div
Fig. 6.23: Light load steady-state operating waveforms. Vbat = 3.0 V, Vdd = 1.5 V, Vout = 0.2 V,
Iload = 1.2 mA.
iL
20 mA/div
vx
1 V/div
Fig. 6.24: Detail of the zero-current NMOS turn-off transition (Vout = 0.2 V).
6.2 An Ultra-Low-Voltage DC-DC Converter 175
efficiencies are acceptable, but are somewhat lower than expected for three primary
reasons. First, the larger than expected inductor value adversely affects the efficiency
of the power train, and increases the average dissipation of the i NMOS comparator
simply because they switch 50% more often. This degrades overall conversion
efficiency nearly equally across the full load range. Second, the extra dissipation
caused by the short-circuit path in the Vref -Vo comparator is not budgeted. According to
circuit simulations, this short-circuit dissipation, present only during evaluation mode,
increases the average power of the comparator by a factor of five at V bat = 3.0 V and
nominal process. Since this comparator dominates the overall quiescent dissipation of
the regulator, light load efficiency is poor. Third, the power-down scheme of the master
bias is disabled on the IC. The bias is attached off-chip, introducing several picofarads
of parasitic capacitance, and increasing the required settling time. This increases
Figure 6.26 shows the mechanisms of loss measured on the regulator at V bat =
3.0 V, Vout = 0.2 V, and P load = 161 µW and 921 µW. The power train losses include the
power transistors with their gate drive, all losses associated with the package, the input
and output capacitors, and the output inductor. Also included in the power train losses
are the power consumption of all digital circuits which run from V bat . The analog power
includes the master bias, voltage reference, Vref -Vo comparator, and i NMOS comparator.
The digital circuits operated from V dd = 1.5 V include only the master control.
Even with a 60% light-load and low-voltage efficiency, it is the key enabler of the low-
80
Vout = 0.2 V
Vout = 0.5 V
75
Efficiency [%]
70
65
60
0 1 2 3 4 5 6
Pload [W] x 10
−3
Power Train
Analog
Digital at Vdd = 1.5 V
250
150
100
50
0
1 2
Fig. 6.26: Mechanisms of loss for Vbat = 3.0 V, Vout = 0.2V, light and heavy load.
6.3 Prototype Dynamic Voltage Scaling DC-DC Converter 177
core and surrounding peripherals are run from a dynamically scaled voltage supply,
Figure 6.27 shows a block diagram of the dynamic DC-DC converter prototype
The DVS loop forces the processor clock frequency, f VCO , to equal the commanded
exploiting the 4 MHz DVS system clock and using low-power digital control
bootstrapped from the converter output, the controller achieves low static power
dissipation which scales together with the load. Pulse-width modulation commands the
quantity of charge delivered during each PFM pulse through the controlled power FET
maintaining acceptable conversion efficiency over the dynamic range of the converter.
6.3 Prototype Dynamic Voltage Scaling DC-DC Converter 178
4 MHz
system clock
Frequency 4 4 Vbat
Detector
Current Comparators:
PMOS limit, NMOS limit
PMOS zero, NMOS zero
Vbat
p_on −
p_off Start-up
Logic +
pwrGD
Soft-start circuits
generates an 8-bit digital representation of the frequency error, f err, every 1 µs. The
loop filter samples f err on the following falling edge of clk4. In the first cycle of Figure
6.28, f err = -1, and the converter idles until the next sampling instant. During this
interval, the processor discharges V dd, causing a corresponding decrease in f VCO . When
the sampled f err > 0, the loop filter translates f err into an update command for the DC-
DC converter. A PFM pulse is initiated by the PMOS power FET, and the power NMOS
functions as a synchronous rectifier, turned off by the NMOS zero current comparator
clk4
clk1
ferr<7:0> -1 +1
update<3:0> 2 1 0
iL
Vdd
vgp
vgn
terms (Figure 6.29). A power FET conduction interval, Ton, is the controlled variable.
f des – f VCO
f err = floor ---------------------------- (Eq 6-14)
1 MHz
RAM M
gain
+ feedforward
fdes + ferr +
Σ Σ Ton
−
fVCO
In (Eq 6-15), the Ton LSB is 250 ns, equal to one cycle of the 4 MHz DVS system clock.
The transfer function is two-sided (Figure 6.30). For f err < 0, Ton < 0 and the
converter removes excess charge from its output capacitor. The PFM pulse is initiated
by the NMOS power FET, T nmos = T on , and the power PMOS is operated as a
synchronous rectifier. For f err > 0, the converter delivers charge to the output via a PFM
Vdd
∆V < 0 ∆V > 0
Vdd
Current limiting is included to protect the power FETs and external filter
elements during large signal tracking transitions. The magnitude of peak positive and
Pulse frequency modulation ensures that the converter switches only when
necessary, conserving power at low output voltage and light load. The pulse-skipping
algorithm is simple: For – 3 ≤ f err < 0 , the converter idles, allowing the processor to
removed from the output according to the PWM algorithm of (Eq 6-15).
600
400
200
Ton [ns]
−400
−600
−10 −8 −6 −4 −2 0 2 4 6 8 10
ferr (1 LSB per MHz)
24), gain LH = gain HL = 1/4 LSB per MHz, feedforward LH = 1 LSB, and feedforward HL
= 0.
6.3.1.3 Start-Up
the digital controller. Figure 6.32 shows a block diagram of the approach.
converter. A simple synchronous PFM scheme, with a constant 500 ns on-time, is used
to ramp the output voltage. Once the output voltage exceeds a weak PMOS V GS ≈ 1.2 V ,
the pwrGD flag is raised, and the DVS controller assumes command of the converter,
initialized with M = 24. When 21 MHz < f VCO < 27 MHz , the TRACK signal falls,
pwrGD TRACK
MHz
50
1 us average VCO
Desired
0
0 0.5 1 1.5 2 2.5
−4
x 10
4
vo [V] 3
1
0 0.5 1 1.5 2 2.5
−4
x 10
2
1
iL [A]
−1
−2
0 0.5 1 1.5 2 2.5
time [sec] −4
x 10
The control system has been verified using matlab simulation. Figure 6.33
shows the simulated tracking performance with V bat = 3.6 V, L = 3.5 µH, and C = 4.7
µF. The large-signal 12 MHz to 90 MHz tracking transition settles within 20 µs.
MHz. The DC-DC converter pulse width and pulse frequency are reduced at the lower
27
MHz
26.5
26
25.5
1 1.5 2 2.5 3 3.5
−5
x 10
1.6
vo [V]
1.55
0.2
iL [A]
0.1
−0.1
1 1.5 2 2.5 3 3.5
−5
x 10
95
MHz
94
93
1 1.05 1.1 1.15 1.2 1.25 1.3 1.35 1.4 1.45 1.5
−4
x 10
3.2
3.15
vo [V]
3.1
3.05
1 1.05 1.1 1.15 1.2 1.25 1.3 1.35 1.4 1.45 1.5
−4
x 10
0.3
0.2
iL [A]
0.1
−0.1
1 1.05 1.1 1.15 1.2 1.25 1.3 1.35 1.4 1.45 1.5
−4
x 10
Fig. 6.34: Simulated regulation waveforms at 26 MHz (top) and 95 MHz (bottom).
6.3 Prototype Dynamic Voltage Scaling DC-DC Converter 185
[Burd98]. Figure 6.35 shows the simulated and modeled performance of the integrated
3.3 V, 100 MIPS operating point, yielding an average full-load current of 135 mA. The
energy per operation scales with voltage as shown in Figure 2.1. This data, together
with the modeled ring oscillator performance of Figure 6.35, is used to generate a curve
Typical processor usage statistics are required to fairly specify the design
objectives of the dynamic DC-DC converter. The frequency of transitions and expected
the relative importance of regulation and tracking metrics. [Pering98] describes four
FAST
120 MODELLED
NOM
100
VCO frequency [MHz]
SLOW
80
60
df/dVdd [MHz/V]
40
20
0
0.5 1 1.5 2 2.5 3 3.5 4 4.5
Vdd [Volts]
120
100
80
Idd [mA]
60
40
20
0
0 10 20 30 40 50 60 70 80 90 100
Throughput [MHz]
DVS voltage scheduling algorithms and reports simulation results on three benchmark
requirements, workload demands, and desired throughput statistics, proving that there
are no “typical usage statistics” for which to design. Thus, to increase its utility in a
made to be a good low-voltage regulator (for improved energy efficiency) and a good
Tracking and regulation metric trade-offs through filter element sizing have
tracking metrics, with acceptable output voltage ripple and low-voltage efficiency, is
the primary design objective. Q L , L, and C are chosen according to (Eq 3-23), (Eq 3-
25), and (Eq 3-26) to sustain full load current in a 4 µs minimum repetition period with
acceptable output voltage ripple. (Eq 4-12), (Eq 3-29), and (Eq 5-12) provide estimates
6.3 Prototype Dynamic Voltage Scaling DC-DC Converter 187
to total losses in the power FETs, conduction loss in the filter elements, and additional
tracking and regulation metrics. A fourfold improvement in tracking time and a sixfold
improvement in tracking energy are expected over previous dynamic DC-DC converters
[Wei96], [Namgoong97], [Kuroda98]. Power train and output voltage ripple losses are
kept below 4% at the low throughput corner. Figure 6.37 shows the charge delivered per
PFM pulse, the PMOS and NMOS conduction intervals, the output voltage ripple, and
Charge delivered per PFM pulse PMOS and NMOS conduction times
550 3
500
2.5
450
400
2
350
QL [nC]
us
300 1.5
250
1 Tpmos
200
150 Tnmos
0.5
100
50 0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Processor throughput [MHz] Processor throughput [MHz]
0.035
0.045
Energy dissipation normalized to energy delivered
0.03 Total
0.04
0.025
0.035
∆V / Vdd
0.02
0.03
0.015
PMOS
0.025
0.01
inductor
0.02
0.005 NMOS
representation of the VCO frequency error averaged over a 1 µs period. The operating
A seven-bit counter clocks rising edges from the VCO output frequency, f VCO .
The reference frequency, f REF = 1 MHz, which is derived from the 4 MHz DVS system
clock, asserts the asynchronous reset of the counter, resetting its output to zero every 1
µs. Just prior to the asynchronous reset, the output of the counter is given by:
f VCO
count(k) = floor ---------------- + remainder(k-1) (Eq 6-16)
1 MHz
where remainder(k-1) is the remainder of the truncation of (Eq 6-16) performed in cycle
(k-1).
This output is latched and subtracted from the 7-bit digital representation of
the desired frequency, M, yielding an 8-bit two’s complement digital error signal:
Operating System M 7 M 7
Loads Desired fCLK
Reg.
(in MHz)
Counter
RST
7 8 To Loop
fREF = 1 MHz
Σ Filter
fVCO
which is proportional to the frequency error, averaged over cycle k, with an LSB of 1
MHz.
However, as illustrated by Figure 6.39, while the error is truncated every 1 µs, the
remainder of the error accumulates in the frequency detector, forcing the average
quantization error to zero. Thus, quantization contributes no DC offset to Vdd and f VCO ,
loading conditions, and therefore, consumes static power. So that its power
consumption scales at lower output voltages, it is operated from the voltage scaled
supply, V dd . The effective capacitance includes a 7-bit counter switching at the VCO
output frequency, a 2-bit clock divider switching at 4 MHz, and a 7-bit register and 8-
bit adder switching at 1 MHz. The average power dissipation is given by:
2 2
P FreqDetect = ( 1.1 pF ) ⋅ f VCO ⋅ V dd + ( 3.6 pF ) ⋅ ( 1 MHz ) ⋅ V dd (Eq 6-18)
cycle (k-1)
remainder = 0.41
cycle k cycle (k+1)
count = 7, remainder = 0.08 count = 6, remainder = 0.59
fREF = 1 MHz
contributing 10 µW at the 5 MHz, 1.05 V operating point, and 1.2 mW at the 100 MHz,
The VCO output is driven from the processor to the dynamic DC-DC converter
IC. If swung rail-to-rail, its power consumption might prove to be the dominant
contributor to overall dissipation in the DVS loop. At the low throughput corner:
2
P VCO = ( 20 pF ) ⋅ ( 1.05 V ) ⋅ ( 5 MHz ) = 110 µW (Eq 6-19)
2
P VCO = ( 20 pF ) ⋅ ( 3.3 V ) ⋅ ( 100 MHz ) = 22 mW (Eq 6-20)
If, instead, the 20 pF of parasitic capacitance is driven by the low-swing I/O transmitter
of Chapter 2 powered by the 200 mV output of the DC-DC converter of Section 6.2, the
2
P VCO = ( 20 pF ) ⋅ ( 0.2 V ) ⋅ ( 5 MHz ) = 4 µW (Eq 6-21)
2
P VCO = ( 20 pF ) ⋅ ( 0.2 V ) ⋅ ( 100 MHz ) = 80 µW (Eq 6-22)
incoming 200 mV signal. A description of the receiver can be found in [Burd98]. Its
2
P receiver = ( 15 µA ) ⋅ V bat + ( 0.9 pF ) ⋅ V dd ⋅ f VCO (Eq 6-23)
6.3 Prototype Dynamic Voltage Scaling DC-DC Converter 191
yielding 45 µW at 5 MHz, 1.05 V, and 1.0 mW at 100 MHz, 3.3 V. The total power
savings effected by the low-swing VCO transceiver is 1.8x at the low throughput
corner, and 20x at the high throughput corner. These numbers include the dissipation in
The loop filter translates f err into an update command for the DC-DC
scheduler. In tracking mode, the converter is capable of slewing its output up and down.
When f err > 0 , the VCO frequency is too low, and the converter is commanded to
deliver charge to the output capacitor. The PMOS device initiates the PFM pulse, T pmos
is the controlled variable, and the NMOS power FET acts as a synchronous rectifier.
When f err < 0 , the VCO frequency is too high, and the converter is commanded to
remove charge from the output capacitor. The NMOS device initiates the PFM pulse,
T nmos is the controlled variable, and the PMOS power FET acts as a synchronous
In regulation mode, the converter can only deliver charge to the output
capacitor, it cannot remove it. When f err ≥ 0 , a PFM pulse is initiated by the power
PMOS device. When f err < 0 , the converter idles and the loop filter continues to
g 16 x 16 M<6:3>
2 SRAM 4
FF clk4
4
Vdd to Vbat
fmag update
P_on
ferr 2’s complement 7 >> + 4
8
REG 8 8
Ton N_on
to sign / mag
ferr<7>
fsgn
>
enable
clk4 TRACK
Fig. 6.40: Loop filter implementation.
Figure 6.40 shows a block diagram of the loop filter implementation. f err
swings at V dd ; all other signals are driven at V bat . The “enable” block implements the
pulse-skipping function, clocking f err on the falling edge of clk4 under the following set
of conditions:
• TRACK is high, or
The 8-bit two’s complement f err is level-shifted to V bat and converted to an 8-bit sign /
controlled.
–g
update = FF + 2 ⋅ f mag (Eq 6-24)
6.3 Prototype Dynamic Voltage Scaling DC-DC Converter 193
which stores Ton in LSB. The loop filter saturates at update = 15, constraining the
maximum on-time to 3.75 µs. Feedforward and gain terms are set as a function of the
four MSBs of the desired frequency, M. Unique values of FF and g are chosen for low-
The “Ton” block negotiates power FET sequencing and converts update into a
The loop filter consumes no static power: It switches only during active PFM
pulses. The energy dissipated per DC-DC converter switching event is data dependent,
2 2
E filter = ( 1.7 pF ) ⋅ V dd + ( 9.2 pF ) ⋅ V bat (Eq 6-26)
which equals 120 pJ (0.2%) at the low throughput corner and 138 pJ (negligible) at the
identical to the one shown in Figure 6.14, for zero-current detection and current
limiting in the power transistors. To conserve quiescent power, strobed biasing and
The PMOS current limit protects the power FETs and external filter elements
during large signal tracking transitions. The peak conducted PMOS current is limited to
comparator
trip point
vx iL
xN x1 +
Poff
vx REF
REF
iL iREF -
IlimP comparator t
Power PMOS Reference generator tCMP
and a known current i REF. The comparator begins to switch when inductor current, i L ,
conducted through the PMOS power FET induces a source-to-drain voltage drop greater
than that induced by i REF flowing through the reference FET. The accuracy of the
is determined primarily by the control on the absolute value of i REF, and the matching
This circuit is activated only during tracking PFM pulses which are initiated
by the PMOS device. It includes a strobed bias network and gated clocks for low-power.
The NMOS current limit is nearly identical to the PMOS current limit of
Figure 6.41. It is activated only during tracking PFM pulses which are initiated by the
1. In the prototype, the current limit may be adjusted with the Ilim_1A pin.
6.3 Prototype Dynamic Voltage Scaling DC-DC Converter 195
NMOS device, and dissipates no static power during regulation mode. Peak negative
application in the low-voltage regulator IC. The NMOS off comparator performs the
identical function in the dynamic DC-DC converter: It commands the turn-off transition
of the NMOS synchronous rectifier when i dsN crosses zero from above.
The equivalent input-referred offset voltage (Vos = 0.5 mV) and delay (t cmp ~
50 ns 2 ) of the comparator are listed in Section 6.2.2.3. In the DVS application, the
0.5 mV 1.05 V
I ε = -------------------- + ( 50 ns ) ⋅ ----------------- = 3.1 mA + 15.0 mA (Eq 6-28)
160 mΩ 3.5 µH
and
0.5 mV 3.3 V
I ε = ------------------ + ( 50 ns ) ⋅ ----------------- = 12.5 mA + 47.1 mA (Eq 6-29)
40 mΩ 3.5 µH
for the low and high throughput operating points. This translates to worst-case energy
adaptive dead-time control (see Section 4.2.3), is used to null the comparator, logic,
and power FET gate-drive delays. Figure 6.42 describes the approach.
comparators. The NMOS off comparator commands the power NMOS turn-off
iREF
REF REF comparator
+ iL
1x matched reference FET Noff trip point
iL
vx pGND
-
1000x NMOS power FET
pGND t
NMOS off comparator
tCMP
Reference generator
vx iREF = N * ILSB
+
up/dn N
+ -1
z
pGND 5 bits
-
Update comparator
Digital integrator
transition. The update comparator monitors the results and adapts the NMOS off trip
where pGND is a Kelvin connection to the power NMOS source terminal. The reference
source, i REF, so that the trip point of the comparator is given by:
W NMOS
i L = i REF ⋅ -------------------- = 1000 ⋅ i REF (Eq 6-31)
W REF
6.3 Prototype Dynamic Voltage Scaling DC-DC Converter 197
Proper adjustment of i REF is ensured by the integral feedback loop. A digital integration
scheme is selected to allow maintenance of state without static power dissipation. The
Gated clocks and strobed biasing are used to eliminate static power. The
comparators are enabled by the power PMOS turn-on − during positive PFM pulses only
− and are disabled 125 ns after NMOS turn-off. The reference generator, with 0 to 62
µA of static current, is enabled 125 ns after NMOS turn-on, and is disabled at NMOS
turn-off. The overall energy dissipated per NMOS off event is given by:
1 2 2
E = --- LI ε + ( 8.0 pF )V bat + V bat ⋅ ( ( 310 µA ) ( T p + T n + 125 ns ) + ( 30 µA )T n ) (Eq 6-32)
2
(Eq 6-32) includes the energy dissipation penalty associated with early or late NMOS
turn-off, and assumes i REF = 30 µA. For V bat = 3.6 V, E = 1.2 nJ (1.7%) at the low
throughput corner. Here, it is interesting to note that the adaptive timing control
The PMOS off comparator is nearly identical to the NMOS off comparator. It
commands the turn-off transition of the PMOS synchronous rectifier when i dsP crosses
zero from below. It includes an adaptive timing control loop to null comparator, logic,
The comparators are enabled by the power NMOS turn-on − during negative
PFM pulses only − and are disabled 125 ns after PMOS turn-off. The bias is never
enabled during regulation mode. Strobed biasing and gated clocks assure that it
The integrated power FETs are binary weighted, with two control bits each for
independent dynamic NMOS and PMOS sizing. The NMOS and PMOS gate-width
LSBs are 10 mm and 20 mm, respectively. The minimum drawn channel length of 0.6
µm is used.
Figure 6.43 shows the power FETs, gate-drive, and dynamic transistor sizing
modules. The FETs are dynamically sized versus requested throughput, M, a-priori,
with appropriate control bits Wp0, Wp1, Wn0, Wn1 stored in RAM. Switching and
gate-drive loss are traded with conduction loss at each operating point according to (Eq
1 2 Rp Rn 2
E diss = --- ⋅ I p ⋅ T p ⋅ -------- + T n ⋅ -------- + V bat ⋅ ( C overhead + W p ⋅ C p + W n ⋅ C n ) (Eq 6-33)
3 W p W n
where subscripts p and n indicate contributions due to PMOS and NMOS power
transistors; I peak is the peak PFM pulse current, found from (Eq 3-19); W is the gate-
width in LSB; T is the conduction time interval, found from (Eq 3-20); R is the
effective channel resistance of an LSB, listed in Table 6.3; and C is the effective
C = C gd + C gs (Eq 6-34)
Figure 6.44 shows the gate-widths and expected energy dissipation for the
prototype IC implementation.
6.3 Prototype Dynamic Voltage Scaling DC-DC Converter 199
Vbat Vbat
Cgdp 2 Cgdp
Cdbp 2 Cdbp
Power PMOS
Cgdn 2 Cgdn
Cdbn 2 Cdbn
Power NMOS Cx
8.98 10.33
p0
20 mm / 0.6 µm
22x 196x
PMOS LSB
8.98 10.33
NMOS LSB
n0 10 mm / 0.6 µm
11x 98x
Wn0 Wp0
n0 p0
3x 1.5x 24x
vgn_in 1x vgp_in 3x
1x 6x 6x
3x 48x
n1 Wp1 p1
Wn1
1 Wpmos
Wnmos
0
0 10 20 30 40 50 60 70 80 90 100
0.025
Epmos
Normalized dissipation
0.02
Enmos
Eoverhead
0.015
0.01
0.005
0
0 10 20 30 40 50 60 70 80 90 100
Throughput [MHz]
Figure 6.45 plots the expected converter efficiency versus throughput at full-
load and at one-quarter-load. The mechanisms of steady-state loss in the DVS system
are summarized in Table 6.4. All losses in the power train, controller, and processor
load are considered. The DVS system is expected to dissipate 138 µW and 3.4 mW of
static power at the low throughput and high throughput corners, with the converter
consuming the majority of the power. Here, the primary mechanisms of dissipation
include the processor VCO, and the VCO receiver, frequency detector, and master bias
of the DC-DC converter. Considering all losses in the processor and converter at full-
load, the system energy per operation is expected to be 0.3 nJ/instruction at 5 MIPS and
0.95
0.9
0.85
Efficiency
0.8
1/4 full load
full load
0.75
0.7
0.65
0.6
0 10 20 30 40 50 60 70 80 90 100
Processor Throughput [MHz]
Fig. 6.45: Expected converter efficiency vs. processor throughput at heavy and medium loads.
PROCESSOR
Low-swing VCO (Eq 6-21) and (Eq 6-22) 4.0 µW 4.0 µW 80.0 µW 80.0 µW
interconnect
The prototype converter was fabricated in a single poly, triple metal CMOS
process through the MOSIS program in August, 1997. Figure 6.46 shows the IC layout,
with die dimensions of 1.68 mm x 3.41 mm. The power section includes 1.6 nF of
integrated bypass capacitance tuned to τ RC = 2.6 ns. Considerable die area is devoted to
the six offset-cancelled comparators, whose offset storage capacitors are implemented
voltage digital and analog supplies are maintained for isolation and power
characterization.
printed circuit board in a through-hole socket. The pinout and pin description are given
in Figure 6.47, and Table 6.5. Table 6.6 estimates the parasitics added in series with the
power train.
6.3 Prototype Dynamic Voltage Scaling DC-DC Converter 203
VX PADS
GATE-DRIVE
pGND PADS
pVDD PADS PMOS NMOS
START-UP
BIAS
68
67
66
65
6
5
4
3
2
1
VX
VX
VX
VX
VX
VX
VX
VX
VX
VX
10 60
pVDD pGND
11 59
pVDD pGND
12 58
pVDD pGND
13 57
pVDD pGND
14 56
pVDD pGND
15 55
pVDD pGND
16 54
RAM_dOUT vgp
17 53
18
RAM_cs DVS Prototype clk4
52
Vbat vgn
19 51
GND 68 LDCC Vbat
20 50
Vdd GND
21 49
GND PORB
22 48
RAM_dIN Vbat
23 47
RAM_clkout GND
24 46
serialM pwrGD
25 45
readM TESTenable
26 44
TRACK Ilim_1A
enextclk
fclk_out
fclk_in
aGND
aGND
extclk
aVdd
GND
ibias
Vref
Vdd
Vfb
30
31
32
33
34
35
36
37
38
39
40
41
Fig. 6.47: IC pinout.
A simplified PCB schematic is given in Figure 6.48. The converter under test
programmable digital CMOS load. The VCO output, routed to the dynamic DC-DC
converter, can be selected to swing at the full V dd rail, or at a reduced 200 mV rail. At
power-on, the internal 16 x 16 SRAM of the converter is loaded with dynamic transistor
sizing codes, and PWM variables gain LH , gain HL , feedforward LH , and feedforward HL
by the NM93CS06LN EEPROM. Frequency requests are loaded serially using the
readM and serialM pins of the converter. Figure 6.49 shows the serial load for f des = 17
MHz.
6.3 Prototype Dynamic Voltage Scaling DC-DC Converter 206
Vbat
Vbat
EEPROM 4 3.5 µH
DVS vx Vdd
Prototype
clk4
4.7 µF
CPU Load
fVCO
LV_fVCO
Idd 4
4 MHz
2
Vbat Vdd
Programmable CPU Load
DAS
Level 0.2 V
or
∼
Converters Transmitter
FPGA
VCO
clk4
readM
serialM
1 0 0 0 1 0 0
requests in the µs to tens of µs time scale, and to regulate with 80% to 90% full-load
efficiencies over the full 5 MHz to 100 MHz dynamic range. The following subsections
iL
TRACK
Vdd
pwrGD
6.3.10.1 Start-Up
Figure 6.50, Figure 6.51, and Figure 6.52 show successful operation of the
captured. Figure 6.51 provides detail of handoff from soft-start to tracking mode and
shows the relative timing of the pwrGD and TRACK flags. Figure 6.52 shows the VCO
and f VCO = 24 MHz, with V bat = 3.3 V and a 0.5 A current limit. Here, the measured
current limit is 360 mA and tracking time is to the order of 10 µs. In Figure 6.53, this
same tracking transition is shown with an expected 1.0 A current limit (measured to be
iL
pwrGD
Vdd
TRACK
Fig. 6.51: Start-up transient, showing relative timing of pwrGD and TRACK.
fVCO
Vdd
TRACK
Fig. 6.52: Start-up transient, showing fVCO = 23.4 MHz when TRACK falls.
6.3 Prototype Dynamic Voltage Scaling DC-DC Converter 209
iL
Vdd
pwrGD
TRACK
In Figure 6.54 and Figure 6.55, full-scale 5 MHz to 100 MHz and 100 MHz to
5 MHz tracking transitions are made with V bat = 6.0 V, at medium load, and with a 1.0
A current limit. The low-to-high tracking time of 23.5 µs is slew limited by the forward
PMOS current limit. The high-to-low tracking transition is slower by design and
measured to be 44.0 µs. The -1.1 A reverse NMOS current limit slew limits the early
portion of the output voltage excursion. The feedback loop intentionally slows the latter
low-to-high frequency transitions at 1/4 full-load and V bat = 6.0 V. Tracking time is
measured from the rising to falling edges of the TRACK signal, yielding the 0% to
f des – 3 MHz points. Tracking energy is estimated for the entire low-to-high-to-low
6.3 Prototype Dynamic Voltage Scaling DC-DC Converter 210
iL
Vdd
TRACK
iL
Vdd
TRACK
tracking cycle from C = 4.7 µF, the measured steady-state dissipation as a function of
Figure 6.56 and Figure 6.57 show regulation at f VCO = 24 MHz, with V bat = 3.3
V, under a large 22 mA load, and a small 1 mA load. The PFM period scales with load
as expected, with the average T = 7 µs and T = 140 µs at heavy and light loads. The
Figure 6.58 and Figure 6.59 show regulation at f VCO = 102 MHz, V dd = 3.78 V
for a commanded M = 100, V bat = 4.0 V, and 1/4 full-load. The output voltage is tightly
Figure 6.60 and Figure 6.61 show regulation at f VCO = 6 MHz, V dd = 1.08 V
for a commanded M = 5, V bat = 4.0 V, and 1/4 full-load. The 29 mV peak-to-peak output
voltage ripple is 2.7% of V dd − slightly larger than anticipated, but still contributing
iL
Vdd (AC)
vgn
vgp
Fig. 6.56: Regulation waveforms at Vdd = 1.47 V, fVCO = 25 MHz, Idd = 22 mA.
iL
Vdd (AC)
vgn
vgp
Fig. 6.57: Regulation waveforms at Vdd = 1.47 V, fVCO = 25 MHz, Idd = 1 mA.
6.3 Prototype Dynamic Voltage Scaling DC-DC Converter 213
Vdd
VCO
iL
Vdd(AC)
vx
Vdd
VCO
iL
Vdd(AC)
vx
92
90
Efficiency [%]
88
86
84
82
80
0 10 20 30 40 50 60 70 80 90 100
Throughput request, M
frequency requests, M. These numbers are generally consistent with expected results,
though they tend to fall off at higher throughput requests. This is attributed to the
higher-than-expected battery voltage, V bat = 5.0 V, necessary to allow the 89 MHz and
100 MHz operating points 3 , and to the additional series resistance of the 68LDCC
Figure 6.63 shows the mechanisms of power dissipation for various loads at
V bat = 3.3 V, V dd = 1.47 V, and f VCO = 25 MHz. The recorded efficiencies are 87%,
85%, and 74% for 22 mA, 11 mA, and 1 mA loads. Power train dissipation, which
includes losses in the power FETs, package, and all external filter elements, dominates
converter losses, even at light load. The VCO receiver and frequency detector are the
50
0
1 2 3
40
Power train
Power Dissipation [mW]
30
Idd = 11 mA Idd = 22 mA
20
10
Idd = 1 mA
0
1 2 3
the NMOS off comparator, is the largest dissipater in the controller at heavy load. All
Figure 6.64 shows the i L , v gn , and v x waveforms for a single PFM pulse at V bat
= 3.3 V, f VCO = 24 MHz. The DC value of V dd is 1.47 V. Figure 6.65 shows detail
around the NMOS power FET turn-off. Here, the power NMOS is turned off at i L < 2
mA, well within the error budget specified in Section 6.3.6.3, and introducing
negligible LI ε 2 loss.
The low-swing VCO transmitter failed on the processor test chip. The low-
swing signal is expected to reach 0 V and 200 mV logic levels, but in Figure 6.66 is
6.3 Prototype Dynamic Voltage Scaling DC-DC Converter 217
iL
vgn
vx
Fig. 6.64: Successful NMOS zero-current turn-off. Vbat = 3.3 V, Vdd = 1.47 V.
iL
vgn
vx
Fig. 6.65: Zoom-in of a successful NMOS zero-current turn-off. Vbat = 3.3 V, Vdd = 1.47 V.
6.3 Prototype Dynamic Voltage Scaling DC-DC Converter 218
Vdd
Low-swing
fVCO
Full-swing
fVCO
seen to be corrupted by noise, and to swing only between ± 50 mV . Since the low-
voltage f VCO never reaches the 100 mV reference voltage level, it cannot be
VCO output is transmitted on the test board, increasing the effective load on the
converter.
6.3.11 Conclusion
demonstrated as a dynamic voltage scaling enabler. Compared with the previous work
tracking metrics.
6.4 A ZVS PWM DC-DC Converter 219
In this section, the design techniques of Chapter 4 are applied to the 6 V to 1.5
effort to validate many of these design techniques, and to demonstrate the viability of
Figure 6.67 shows a block diagram of the chip. The IC is operated as an open-
voltage switching transitions are guaranteed from zero to full load through the
adjustable dead-time control blocks, labelled τ DHL and τ DLH . Power transistors and
drivers are designed to minimize total power transistor losses; in this case, gate and
The power train of the low-voltage buck circuit, with device sizes and external
component values, is shown in Figure 6.68. All active devices are integrated on a single
die and fabricated in a standard 1.2 µm single-poly double-metal CMOS process. The
Vbat = 6 V
driver
1 MHz
τDHL vgp iL
25% duty
osc vx Vo = 1.5 V
PWM L
driver
τDLH vgn Cx C LOAD
Fig. 6.67: Block diagram of the 6 V to 1.5 V, 500 mA prototype buck converter.
6.4 A ZVS PWM DC-DC Converter 220
+
vgp
10.2 cm
-------------------
0.9 µm
675 nH
6V
+
vgn
10.5 cm
------------------- 4 nF 20 µF 1.5 V
0.9 µm
- -
circuit exhibits nearly ideal ZVS using an adjustable dead-time control scheme similar
Figure 6.69 shows the ideal periodic steady-state waveforms. The inverter
25%, and an operating frequency of f s = 1 MHz which allows a compact, yet highly
efficient converter. The inductor current reverses to allow ZVS transitions of both
6V 1 µs
vx
0
0.5 A 1.66 A
iL 0
6V
vgp
0
6V
vgn
0
100 ns 25 ns
power transistors, eliminating the loss associated with Miller charge and all stray
capacitance attached to v x .
Because the inverter node transition intervals are designed to be small relative
to the switching period, i L is assumed triangular with peak negative and positive values
I o – ∆I ⁄ 2 and I o + ∆I ⁄ 2 which are constant over the entire dead-time. The ratio of
inverter node transition times is given by the ratio of currents available for each
commutation:
τ xLH ∆I ⁄ 2 + I
- = -----------------------o-
----------- (Eq 6-36)
τ xHL ∆I ⁄ 2 – I o
and approaches unity for large inductor current ripple. In (Eq 6-36), τ x indicates an
inverter node transition time, with subscripts LH and HL denoting low-to-high and
high-to-low transitions, respectively, I o is the average load current, and ∆I is the peak-
current ripple of
∆I 5
------ = --- ⋅ I o = 833.3 mA (Eq 6-37)
2 3
Vo ⋅ ( 1 – D )
L = ----------------------------- = 675 nH (Eq 6-38)
f s ⋅ ∆I
Vo ⋅ ( 1 – D )
C = --------------------------------- = 13.9 µF (Eq 6-39)
2
8 ⋅ L ⋅ ∆V ⋅ f s
and C = 20 µF is selected.
( ∆I ⁄ 2 – I o ) ⋅ τ xLH
C x = --------------------------------------------- = 5.6 nF (Eq 6-40)
V bat
a reasonable value.
on-chip one-shots (the τ DLH and τ DHL blocks in Figure 6.67). External potentiometers
allow manual trimming of these delays to estimate the power savings effected by ZVS.
Figure 6.70 shows two measured non-ideal ZVS high-to-low inverter node
transitions. In Figure 6.70a, τ xHL > τ DHL , so that the NMOS turns on early, discharging
C x through its resistive channel and introducing C x V bat 2 loss. In steady-state, if both
power transistors fully (dis)charge C x over the full potential, V bat , nearly 200 mW of
additional power dissipation is introduced. In Figure 6.70b, τ xHL < τ DHL , so that the
NMOS turns on late, inducing greater than 30 ns of body diode conduction and reverse
recovery loss. With 30 ns of high-side and low-side body diode conduction each cycle,
6.71. Here, the NMOS device is turned on approximately when v x = 0, introducing little
The power transistors are sized according to (Eq 4-11) to minimize their total
losses in periodic steady-state at full load. The minimum effective channel length, L eff
= 0.6 µm, is used. Device parameters R 0 and Q g0 = E g0 ⁄ V bat , which represent the
(a) (b)
Fig. 6.70: Non-ideal ZVS transitions: (a) The NMOS is turned on early. (b) The NMOS is turned
on late. The upper trace is vgn, the lower trace is vx, the vertical scale is 2 V/div, and the horizontal
scale is 20 ns/div.
Fig. 6.71: Ideal ZVS high-to-low inverter node transition. The upper trace is vgn, the lower trace is
vx, the vertical scale is 2 V/div, and the horizontal scale is 20 ns/div.
6.4 A ZVS PWM DC-DC Converter 224
effective channel resistance and gate charge of a minimum gate-width device, are found
extracted layout of large geometry FETs to W 0 = 0.6 µm, the minimum feature size in
the 1.2 µm process. Plugging C g0 = Q g0 ⁄ V bat and all necessary application- and
Approximate power transistor gate-widths are found by substituting this estimate and
tapering factor between successive inverters, u, and the number of inverters in the
chain, N, with (Eq 4-32) and (Eq 4-33). Iteration using circuit simulation on extracted
layout is beneficial to refine the design. From (Eq 4-12), total FET losses at full load
PMOS NMOS
R0 23.7 kΩ 6.2 kΩ
Buffering, N 4 4
The circuit as presented in [Stratakos94] uses the full battery input voltage to
drive the gates of the power transistors. To gain a modest improvement in efficiency,
bootstrap the gate-drive from the 1.5 V output of the converter. With V g = Vo = 1.5 V,
total FET losses may be reduced from 5.9% to roughly 4% at full load, but at the
6.4 A ZVS PWM DC-DC Converter 225
PMOS
τD control
oscillator
gate-drive
NMOS
process through the MOSIS program. The circuit successfully delivers 750 mW at 1.5 V
from a 6 V supply. Figure 6.73 shows the measured steady-state v gp , v gn, i L , and v x
can be observed.
transistor gate and conduction losses are balanced and predicted well by theory and
anticipated. This can be attributed to several factors. First, due to an undetected layout
comparatively large static currents are required to obtain the desired dead-times. Thus,
the power consumption of the ADTC circuitry is greater than an order of magnitude
larger than anticipated, comprising nearly 30% of the overall loss. Second, throughout
the design, efficiency is traded for testability: a number of intermediate signals are
6.4 A ZVS PWM DC-DC Converter 226
vgn
vgp
iL
vx
Fig. 6.73: Measured steady-state waveforms: vgn, vgp, iL, vx (top to bottom). The horizontal scale
is 200 ns/div. The vertical scale is 2 V/div for the voltage waveforms, and 1 A/div for the inductor
current waveform.
severe penalty in dynamic power consumption. For example, the dissipation of the
major component of loss is accredited to the package and test board. The IC is
board. Series resistance in the V bat , ground, and v x lines contribute a total of 47.3 mW
of loss (28% of the total loss), and the stray inductance in the loop formed by the input
(10% of the total loss). Dissipation from these mechanisms can be significantly reduced
by using a smaller surface mount package soldered directly to a printed circuit board.
PMOS NMOS
PMOS NMOS
The results measured on the prototype indicate that in this circuit, on-chip
losses (including those in the power transistors, drivers and control circuits) can be kept
below 8% at full load. The design approach presented in Chapter 4 is evidently viable
for realizing a high efficiency and compact power converter for portable battery-
operated applications. This circuit requires only one custom IC, three small ceramic
chip capacitors, and one small inductor, and is capable of achieving efficiencies above
90%.
7.1 Conclusions 228
Chapter 7
Conclusions
7.1 Conclusions
CMOS design have assumed that the supply voltage is a free variable and can be set to
any arbitrarily low level with little penalty. This thesis has introduced the DC-DC
processors, enabled by small and highly integrated DC-DC converters custom designed
for their individual loads, can provide up to an order of magnitude reduction in overall
scaling applied to ultra-low-swing bus transmitters is used to reduce the power of high-
output DC-DC converter. Dynamic voltage scaling (DVS) is proposed to trade general-
for the key low-power enabler, called the dynamic DC-DC converter, have been
techniques have been presented to reduce the size, cost, and energy dissipation of low-
voltage DC-DC converters. Measured results on three prototype DC-DC converter ICs
have successfully demonstrated these design techniques. The approach presented in this
thesis is evidently viable for realizing compact and highly efficient DC-DC converters
low-voltage and low-power enablers. This has included the development and
utility of DC-DC converters in nearly any portable electronic application. Several key
• Developed a series of design techniques which decrease the size, cost, and energy
dissipation of low-voltage DC-DC converters. These include new ideas, such as:
reported output voltage and power levels: Greater than 70% efficiency at 0.2 V
build.
This thesis has provided the groundwork for a variety of continuing research
converters, or on the portable electronic systems whose battery run-time they are
intended to improve.
single IC, or integration of the DC-DC converter together with its individual digital
CMOS load would offer the smallest size power delivery system. Design of DC-DC
converters as drop-in macros, similar to DSP cores, could be pursued as the next true
References
[Burd96] T. Burd, “Processor Design for Portable Systems”, Journal of VLSI Signal
Processing Systems, vol. 13, pages 203-221, 1996.
232
[Chew91] W. Chew and P. Evans, “High Frequency Inductor Design Concepts”, 22nd
Annual Power Electronics Specialists Conference, pages 673-678, June 1991.
[Dancy97a] A. Dancy and A. Chandrakasan, “Ultra Low Power Control Circuits for
PWM Converters”, IEEE Power Electronics Specialists Conference., pages 21-
27, 1997.
[Frank97] D. Frank, et. al., “Supply and Threshold Voltage Optimization for Low Power
Design”, Proc. 1997 International Symposium on Low Power Electronics and
Design, pages 317-322.
[Hiraki94] M. Hiraki, et. al., “Data-Dependent Logic Swing Internal Bus Architecture
for Ultra-Low-Power LSIs”, Symposium on VLSI Circuits, pages 29-30, 1994.
[Igarashi97] M. Igarashi, et. al., “A Low-Power Design Method Using Multiple Supply
Voltage”, Proc. 1997 International Symposium on Low Power Electronics and
Design, pages 36-41.
[Intel97] Pentium Pro Processor Power Distribution Guidelines, Intel Application Note
AP-523.
[Kunii95] S. Kunii, “Means of Realizing Long Battery Life in Portable PCs”, 1995
IEEE Symposium on Low Power Electronics, pages 20-23.
[Kuroda98] T. Kuroda, et. al., “Variable Supply-Voltage Scheme for Low-Power High-
Speed CMOS Digital Design”, IEEE Journal of Solid-State Circuits, vol. 33,
no. 3, pages 454-463, March, 1998.
[Lau97] W. Lau and S. Sanders, “An Integrated Controller for a High Frequency Buck
Converter”, IEEE Power Electronics Specialists Conference, pages 246-254,
1997.
[Liu93] D. Liu and C. Svensson, “Trading Speed for Low Power by Choice of Supply
and Threshold Voltages”, IEEE Journal of Solid-State Circuits, 28(1), Jan.
1993.
[Locascio93] J. Locascio and W. Cho, “New Controllers for Battery Systems Increase
Systems Efficiency”, in Power Quality USA, 1993.
[Lynn95] L. Lynn, Low Power Analog Circuits for an All CMOS Integrated CDMA
Receiver, Master’s Thesis, University of California at Berkeley, 1995.
[Nakkagone93] Y. Nakkagone, et. al., “Sub-1-V Swing Internal Bus Architecture for
Future Low-Power ULSI’s”, IEEE Journal of Solid-State Circuits, 28(4), pages
414-419, Apr. 1993.
[Shin94] H. Shin, et. al, “Low-Dropout On-Chip Voltage Regulator for Low-Power
Circuits”, 1994 IEEE Symposium on Low Power Electronics, pages 76-77.
235
[Travis97] B. Travis, Sr. Technical Editor, “Powering the Big Microprocessors”, EDN,
pages 31-43, Aug. 15, 1997.
[Usami95] K. Usami and M. Horowitz, “Clustered Voltage Scaling Technique for Low-
Power Design”, Proc. 1995 International Symposium on Low Power Design,
pages 3-8.
[Wang97] T. Wang, X. Zhou, and F. Lee, “A Low Voltage High Efficiency and High
Power Density DC/DC Converter”, IEEE Power Electronics Specialists
Conference., pages 240-245, 1997.
[Wei96] G. Wei and M. Horowitz, “A Low Power Switching Supply for Self-Clocked
Systems”, 1996 International Symposium on Low Power Electronics and
Design, pages 313-317.
[Wu97] A. Wu, S. Sanders, and R. Rossetti, “Active clamp circuits for switchmode
regulators supplying microprocessor loads”, Proc. IEEE Power Electronics
Specialists Conference, volume 2, pages 1179-1185, 1997.
[Yin92] G. Yin, F. Op’t Eynde, W. Sansen, “A High-Speed CMOS Comparator with 8-b
Resolution”, IEEE Journal of Solid-State Circuits, 27(2): 208-211, Feb 1992.