1. Introduction
In CPU and MCU applications, there is a consistent pursuit of high-frequency point-of-load (PoL) power supplies [
1,
2,
3]. The switching frequency has been increasing in the past few decades [
4]. This miniaturizes switches and inductors that primarily dominate the system size and weight. As a result, the integration level has evolved into a 3D stacked power supply (3D-SPS, discrete switch and discrete inductor), power supply in package (PSiP, integrated switch and discrete inductor), and power supply on chip (PwrSoC, integrated switch and integrated inductor) [
5,
6,
7]. These integration levels differentiate in switching frequency and power rate and applications, as shown in
Table 1.
The 3D-SPS adopts discrete switches and inductors, which allows for high power conversion (up to 100 W) [
8,
9,
10]. It carries out a 3D stacking structure on the PCB, and usually places magnetic components above or below the entire PCB board. With relative high inductance, the 3D-SPS approach can operate at a switching frequency below 10 MHz. The reduced switching frequency benefits the overall efficiency, since the switching loss is proportional to the switching frequency.
The PSiP integrates all components in one package, which achieves single chip power conversion [
7,
11]. Typically, it integrates power switches, drivers and control modules on a single die, whereas discrete inductors with low profile are packaged into the chip. Owing to the reduced inductance, the PSiP approach usually operates at a frequency above 10 MHz that leads to high switching loss and reduced efficiency. Due to limited power devices and heat dissipation, the power range is restricted to under 10 W.
The PwrSoC integrates all components on a single die, which achieves the highest integration level [
12,
13,
14]. The integrated on-chip power supply can be placed as near to the load as possible, which reduces the transmission loss in power lines [
15]. However, with a low on-chip inductance, the switching frequency usually increases above 100 MHz to reduce ripples. As a result, power loss and heat dissipation are hard to optimize and that limits the power rate to below 2 W.
In comparison to 3D-SPS, the PSiP saves the interconnecting area of the control chip, which drives chip and power transistors in 3D-SPS. In comparison to PwrSoC, the PSiP generally adopts the standard CMOS process and does not require the CMOS process with power inductors, which is more mature and compatible with technology. Therefore, the PSiP power supply is more widely applied in communication, server automatic systems and so on.
For all integration levels, switching frequency and power levels are usually contradictory, and the major challenges toward high frequency are power losses of switches and inductors, as shown in
Figure 1.
Among the major challenges, the power switch dominates conduction and switching losses. Conduction and switching losses are primarily determined by the on-resistance value (
Ron) and total gate charge value (
QG), respectively. Figure-of-merits (FOM =
Ron ×
QG) of the Si vertical trench, Si lateral trench and GaN lateral trench are shown in
Figure 2 [
16,
17,
18]. Due to the wider bandgap, higher electron mobility and electron velocity of GaN HEMT, its FOM is several times lower than that of Si MOSFET. These material characteristics result in lower power losses in GaN HEMT in high-frequency applications. However, GaN HEMTs are usually discretely packaged due to unique fabrication processes, which are preferable in 3D-SPS [
19,
20,
21]. For low-power and high-integrated applications, the Si lateral diffused MOSFET (LDMOSFET) is usually applied in PSiP and PwrSoC because of compatible fabrication processes [
22].
The power inductor dominates core and coil losses and primarily determines integration levels. Although the magnetic core helps increase the inductance, the core loss increases rapidly with frequency that harms the efficiency. For magnetic cores, their magnetic permeability,
μi, of soft magnetic material decreases dramatically when the operating frequency reaches a critical value, which causes the dramatic decrement of the inductance value. In TDK Mn-Zn ferrites, MAGNETICS nanomaterial and alloy powders, their initial relative magnetic permeability,
μi, is shown in
Figure 3 [
23,
24,
25].
Figure 3 shows that the permeability of magnetic materials decreases rapidly above 1 MHz. As the switching frequency increases, high-frequency inductors evolve from magnetic-core inductors to air-core inductors [
26,
27,
28]. Without magnetic cores, air-core inductors do not suffer core loss and have good linearity to frequency. The inductance value decreases dramatically due to the removal of the magnetic core. In the literature and products, the typical inductance values of solenoid and planar spiral inductors are given in
Table 2. All the investigated inductance values are less than 500 nH, which pushes the PoL power supply to operate at a frequency above 1 MHz.
To address the above challenges, this paper analyzes integrated Si LDMOSFTs and air-core inductors in PSiPs. In terms of the integrated Si LDMOSFET, with consideration of parasitic resistors and capacitors, the size and parallel quantity optimization of integrated Si LDMOSFET are provided to reduce the switching and conduction losses. In terms of air-core inductors, with consideration of skin effect, quality factors of various air-core inductors are calculated for the winding structure optimization to reduce coil loss. Furthermore, combining the above optimizations, this paper predicts the efficiency boundaries of PSiP approaches based on power losses of switches and inductors.
This paper is organized as follows.
Section 2 analyzes power losses of an integrated Si LDMOSFET to provide a size and parallel quantity optimization.
Section 3 calculates the quality factors of solenoid inductors and planar spiral inductors to optimize winding structures.
Section 4 predicts the efficiency boundaries of PSiP.
Section 5 concludes this paper.
2. Power Loss Analysis of Integrated Si LDMOSFET
In PSiPs, integrated Si MOSFETs have the most severe heat dissipation and the easiest thermal breakdown of all power devices. For PSiP, power loss caused by high frequency is a great challenge in terms of designing power switches. Power losses of integrated Si LDMOSFETs mainly include switching loss, driving loss, conduction loss and other power losses [
35,
36]. The equivalent model and switching process of an integrated Si LDMOSFET is shown in
Figure 4. In
Figure 4b,c, the gate voltage
VGS exits the miller platform due to the inductor in the output filter.
Switching loss: At the period
t1~
t3 in turn-on and turn-off processes, the triangular overlapping area of drain-source current and voltage are switching losses. Therefore, according to the overlapping area, the switching loss is given by
Driving loss: At the period
t0~
t5 in turn-on and turn-off processes, gate voltage,
VGS, rises or falls via the charge,
CGS, or discharge,
CGD. The stored energies of
CGS and
CGD are dissipated by the driving resistor,
RG, which depends on the gate charge,
QG. The driving loss is given by
Conduction loss: In on-state, the Si LDMOSFET is equivalent to the on-resistor,
Ron, dissipating the energy. Conduction loss is given by
Other power losses: Other power losses mainly include output capacitance power loss and body diode power loss. The output capacitor stores the energy in off-state and releases the energy via the conducting channel in on-state. Based on the stored and released energies in a switching cycle, the output capacitance power loss is given by
The body diode power loss includes forward conduction loss and reverse recovery loss. The forward conduction loss is caused by the forward body diode current during the dead time. It is given by
where
VF and
td are the forward conduction voltage and forward conduction time of the body diode.
The reverse recovery loss is caused by the body diode reverse recovery after carrying the forward current. It is expressed as
where
VDR and
Qrr are the reverse recovery voltage and reverse recovery charge of the body diode.
Combining (1)–(6), the total power loss of
M switches in parallel is given as
where
X1,
X2,
X3,
X4,
X5 and
X6 represent coefficients in total power loss that do not depend on switching frequency and output current.
Based on the capacitance value per unit length and area,
CISS,
COSS and
CRSS of the integrated Si LDMOSFET are given by [
37]
Based on I–V characteristic of the integrated Si LDMOSFET in the deep linear area, the on-resistance value is given by
Based on (8) and (9), (7) can be derived as
where coefficients
Y1,
Y2,
Y3 and
Y4 are obtained from
X1,
X2,
X3,
X5 and
X6 by excluding the channel width,
WM.
In (10), the power loss of the switch is determined by switching frequency, fs, output current, Io, switch size, WM, and parallel quantity, M. As usual, fs and Io are determined by the power supply requirements. In order to reduce power loss, WM and M are optimized as follows.
M = 1: According to (10), taking
Vin = 5 V and the 350 nm process as an example, the power loss is plotted with the integrated Si LDMOSFET size in
Figure 5. It is obvious that there are different optimal sizes to minimize power loss at different switching frequencies and output currents [
38]. This optimal size can be obtained by the simulation scan.
M ≠ 1: In fact, switch size is limited by the process. Therefore, the power switch is composed of multiple units in parallel. According to (10), the power loss is plotted with the output current in
Figure 6. The size of a single unit is optimized at
fs = 10 MHz and
Io = 0.25 A, according to
Figure 5. In
Figure 6, comparing
M = 4 to all load current ranges, the parallel quantity of power switch changing with drain-source current is more beneficial to improving overall efficiency [
39].
3. Quality Factor Analysis of Air-Core Inductors
In PSiPs, the power loss of air-core inductors can be increased due to high-frequency effects. For megahertz PoL applications, power losses of air-core inductors are related to the inductance current, geometric parameters of coil-wire, coil material, skin effect, eddy effect, proximity effect, fringe effect, etc. Among many factors, the geometric parameters and skin effect are highly related to the switching frequency. To calculate the power loss of the air-core inductors, the inductance current is expanded into sinusoidal waves of various frequencies. Their frequencies and amplitudes are
fs, 2
fs, 3
fs, … and
IL1,
IL2,
IL3, … respectively. Upon consideration of the geometric parameters and skin effect, in [
40], the power loss of air-core inductors can be estimated by
In (11), the inductance power loss consists of the DC loss (
Pind,DC) and AC loss (
Pind,AC). Smaller
RAC at the same inductance value is more beneficial for efficiency, especially in reducing high-frequency power loss. In order to simplify the analysis, the quality factor,
Q, is optimized to decrease the AC power loss, as given by
where higher
Q means a lower
AC resistance value at the same inductance value.
According to the profile, air-core inductors are classified into planar spiral inductors and solenoid inductors, as shown in
Figure 7, which are discussed in the following sections.
3.1. Winding Structure Optimization of Planar Spiral Inductors
For a planar spiral inductor, the outer diameter,
D, inner diameter,
d, coil turn,
N, coil-wire thickness and width determine its profile and inductance value. Inductance values of planar spiral inductors with different winding structures are uniformly expressed as follows [
40]:
where
D and
d are the outer diameter and inner diameter, respectively,
T is the ratio of difference and sum between outer and inner diameters, and
N is the coil turn. The coefficients
q1,
q2,
q3 and
q4 of different winding structures are given in
Table 3.
Considering skin effect and coil length, the
nth harmonic
AC resistance values of planar spiral inductors are given by
where the coefficient,
p, of different winding structures are as given in
Table 3.
Combining (13) and (14), the quality factor of planar spiral inductors is derived as
According to (15), the quality factor depends on the cross-section perimeter,
φ, coil turn,
N, and
T. For
φ and
N, the quality factor increases monotonically with them. For
T, the function-related
T in
Q is defined as
Based on (16), the relationships between
F and
T are plotted in
Figure 8 (
T < 1). According to
Figure 8, at the same
D (area constraint), increasing
d is beneficial for improving quality factor,
Q.
In general, the outer diameter
D is constrained by the inductance value. In order to improve the quality factor,
d,
N and
φ should increase as much as possible. However, the above measures are contradictory in the winding structures of equal width. Therefore, in [
41], wide outside winding and the narrow inside winding are applied to improve the quality factor, as shown in
Figure 9. The quality factor of the proposed winding structure is calculated using the Greenhouse algorithm. Each turn of the winding structure is approximated as a polygon. The inductance value is the sum of self inductances and mutual inductances of all turns. The AC resistance value is the sum of AC resistance values for all turns. Therefore, the quality factor can be calculated based on the inductance and AC resistance values.
3.2. Winding Structure Optimization of Solenoid Inductors
For a tightly winding solenoid inductor, the winding diameter,
ϕ, coil turn,
N, and coil section diameter,
dCu, determine its profile and inductance value. When the winding structure is cylinder, the inductance value of the solenoid inductor is given by [
40] the following:
where Nagaoka’s coefficient,
k, depends on
ϕ/(
NdCu), as given in
Table 4.
The length of the solenoid inductor is given by
Combining (14), (17) and (18), the quality factor of the solenoid inductor is derived as
According to (19), the quality factor depends on
N,
dCu and
ϕ. For
N and
dCu, the quality factor increases monotonically with them. For
ϕ, the function -related
ϕ in
Q is defined as
Based on
Table 4 and (20), the approximation relationship between
G and
ϕ/(
NdCu) is plotted in
Figure 10. According to
Figure 10, at the same
NdCu, increasing
ϕ is beneficial for improving quality factor,
Q, which causes the large volume. Unlike in planar spiral inductors, there is no optimized structure to improve the quality factor.
4. Efficiency Boundary Prediction of PSiP
Power losses of PSiPs mainly include power losses of switches and inductors. Only considering power losses of inductors and power switches, the PoL power supply efficiency is expressed as
where
Po,
Pind and
Pswitch are the output power, power losses of inductors and power switches, respectively.
In order to calculate the power losses of switches and inductors with the consideration of switching frequencies and power levels, Vin = 5 V, Vo = 1.8 V, fs = 1~100 MHz, △iL/IL0 = 0.4 and Po = 0.1~10 W are used in calculations.
For the power loss of integrated Si LDMOSFETs, optimized sizes of integrated Si LDMOSFETs at 350 nm, 180 nm and 90 nm processes are calculated at different switching frequencies and output currents according to (7) and (10). The parameters of optimized integrated Si MOSFETs are given in
Table 5. As the switching frequency increases, the
RDS of integrated Si MOSFET increases, and
CISS,
CRSS and
COSS decrease, which trades off power losses among switching loss, driving loss and conduction loss. Furthermore, based on the parameters of Si LDMOSFETs, the power losses of Si LDMOSFETs,
Pswitch, at 350 nm, 180 nm and 90 nm are shown in
Figure 11a. Furthermore, the efficiency boundary,
Pswitch/
Po, is calculated as shown in
Figure 11b.
In
Figure 11a, the optimized power loss of the integrated Si LDMOSFET is proportional to
fs1/2 instead of
fs, which shows the potential for high-frequency switches. As the switching frequency increases, the power loss of the integrated Si LDMOSFET increases at a rate of
fs1/2. As the power level increases, the power loss of the integrated Si LDMOSFET increases at a rate of
Po. In addition, the power losses of the integrated LDMOSFEET under more advanced processes are reduced due to reductions of the parasitic capacitors and resistors.
For the power loss of air-core inductors, the required inductance value is calculated according to the switching frequency and ripple ratio of the inductance current. Then, with consideration of the optimized winding structures, the geometric parameters (
D,
d,
ϕ) are estimated based on the inductance value according to (13) and (17). Furthermore, based on (14) and (18), the coil length is calculated according to inductance geometric shapes. Finally, based on (11), the power losses of air-core inductors,
Pind, with winding structure optimizations are shown in
Figure 12a. Furthermore, the efficiency boundary,
Pind/
Po, is calculated as shown in
Figure 12b. The following assumptions are introduced into the calculations according to common engineering values:
For the tightly winding solenoid inductor, its wire width and coil turn are set as 1 mm and 10.
For the planar spiral inductor, its thickness and coil turn are set as 100 μm and 10. Since it has a wide outside winding and narrow inside winding structure, its width is set as 1~2 mm.
In
Figure 12a, as the switching frequency increases, the coil-wire length,
l, and skin depth,
δ, both decrease. Based on calculations, since the effect of coil-wire length reduction on power loss is greater than that of skin effect at 1~100 MHz, the power losses of air-core inductors decrease as the switching frequency increases. As the power level increases, the inductance current,
IL, and ripple, △
iL, increase. The increase of △
iL results in reductions of inductance,
L, and coil-wire length,
l. Therefore, the power losses of air-core inductors are at the minimum value as the power level increases.
Combined with power loss calculations,
Pswitch/
Po and
Pind/
Po are shown in
Figure 11b and
Figure 12b. It can be seen that the main power losses at low and high frequencies are from switches and inductors, respectively. Therefore, the directions of further reduction of power losses at low frequency and high frequency are different. At low frequency, the sizes of air-core inductors are slightly increased to reduce their DC and AC resistance values. At high frequency, advanced processes are applied to reduce the parasitic capacitance values of Si LDMOSFETs. According to the above analysis, for reducing power loss to improve switching frequency, PSiPs need to optimize the size of the integrated Si MOSFET in advanced processes and optimize the winding structure of air-core inductors. At a high enough switching frequency, air-core inductors are integrated in the chip, which represents the trend from PSiP to PwrSoC. Reducing power loss is also a heat dissipation requirement for the highly integrated PwrSoC package.
According to
Figure 11b and
Figure 12b, the efficiency boundaries of PSiPs are predicted and verified in
Figure 13 [
42,
43,
44,
45,
46,
47,
48,
49,
50,
51,
52,
53,
54,
55,
56,
57,
58,
59,
60,
61]. In order to eliminate the effect of input and output voltages on efficiency, the efficiencies in the literature are normalized according to the following:
where
ηst and
η are efficiencies with and without normalization.
In
Figure 13, the efficiencies found in state-of-the-art research are lower than those of the predicted efficiency boundary, and most of them are close to the boundary, which verifies the prediction based on the power loss analysis. The predicted errors are mainly from power losses of the equivalent series resistor and controller. The above optimized measures reduce power losses and improve efficiency, which provides roadmaps for achieving high-frequency PSiPs.