RTU1A-1

A Continuous-Mode 23.5-41GHz Hybrid Class-F/F-1 Power

Amplifier with 46% Peak PAE for 5G Massive MIMO
Applications
Tso-Wei Li #1, Hua Wang#2
#
Georgia Tech Electronics and Micro-System, Georgia Institute of Technology, USA
1
tsoweili@gatech.edu, 2hua.wang@ece.gatech.edu

Abstract—This paper presents a continuous-mode hybrid only 24.2%. In [3], the authors present a PA with 46.4%
Class-F and inverse Class-F (i.e., Class-F/F-1) power amplifier PAE, but the BW1dB is only 26.7%. In [10], the authors
(PA) to achieve high efficiency and wide bandwidth covering present a continuous-mode Class-F-1 PA that covers 19 to
the potential 5G bands of 28, 37, and 39GHz. The proposed
continuous-mode harmonically tuned output network 29.5GHz (BW1dB=43.3%). To address this challenge, we
provides the proper harmonic impedance terminations for the propose in this work a continuous-mode hybrid Class-F and
continuous Class-F and Class-F-1 operation modes at lower inverse Class-F (i.e., Class-F/F-1) PA to achieve high PAE
and higher frequency bands, respectively. Moreover, both and instantaneous broadband operation to span multiple
modes present high efficiency and deliver almost the same mm-Wave 5G bands of 28, 37, and 39GHz, similar to what
saturated power Psat with a Psat variation less-than 0.1dB. was reported in [1]. Also, the proposed PA does not require
The proposed PA is implemented in a 45nm CMOS SOI
process, achieving 46% peak PAE, 54.3% Psat output power
any tunable elements or switches, which allows for ultra-
1dB bandwidth (23.5GHz to 41GHz), and 51% small-signal compact chip size and simple implementation.
3dB bandwidth (25.9GHz to 43.7GHz). Also, it achieves 43.4%
PAE and 18.6dBm Psat at 27GHz, 40.2% PAE and 18.6dBm II. POWER AMPLIFIER DEIGN
Psat at 37GHz, and 41.2% PAE and 18.5dBm Psat at 39GHz,
respectively, outperforming reported mm-Wave silicon PAs at Figure 1a shows the schematic of our proposed single-
stage differential continuous-mode hybrid Class-F/F-1 PA.
similar frequencies. This is the first demonstrated continuous-
mode hybrid Class-F/F-1 PA in CMOS to cover the 28, 37, and The proposed PA is realized using a cascode topology with
39GHz mm-Wave 5G bands. identical sizes (W/L=6×30μm/40nm) for M1, M2, M3, and
Index Terms —Class-F, inverse Class-F, continuous mode, M4 that are biased at VG=0.3V, Vcas=1.3V, and VDD=2V. It
fifth-generation (5G), millimeter-wave, power amplifier. utilizes neutralization capacitors Cn=55fF to improve
power gain, reverse isolation, and stability. The input
I. INTRODUCTION matching network is implemented by a 1:1 transformer with
parallel capacitors Cin=160fF and a parallel resistor
Multiple mm-Wave frequency bands have been proposed Rg=170Ω. Both transformers of the input and output
for 5G communication in different countries and areas, matching network utilize the top two metal layers (i.e.,
including 27.5-28.35GHz and 37-40GHz for US, 24.25- 4.1μm aluminum and 3.9μm copper).
27.5GHz and 31.8-33.4GHz for Europe, and 24.25- Existing continuous-mode PA designs require multiple
27.5GHz and 37-42.5GHz for China [1]. An ultra- inductors and capacitors to realize 2nd- and/or 3rd-order
broadband mm-Wave transmitter (TX) that can cover all harmonic impedance terminations and tuning, which
these potential 5G bands will enable frequency diversity substantially increases passive loss, design complexity, and
and 5G international roaming and support massive MIMO chip size [2]-[5]. To overcome these issues, we propose a
systems with ultra-compact elements by eliminating the continuous-mode harmonically-tuned output network,
need for assembling several single-band TXs. In addition, which only occupies an ultra-compact single transformer
the energy efficiency of a wireless TX is often dominated foot-print without additional tunable components or
by its power amplifier (PA), and massive MIMO systems switches (Fig. 1b), making the demonstrated PA
demand high-efficiency PA/TX to meet thermal handling particularly suitable for multi-band 5G massive MIMOs.
The proposed PA output network exploits and enhances
requirements or mobile applications.
the parasitic elements in a physical on-chip transformer to
However, based on reported mm-Wave PA designs [2]-
achieve continuous-mode harmonic tuning in both
[10], it is exceedingly challenging to achieve over 50% Psat
differential- and common-modes with substantial network
power 1dB bandwidth (BW1dB) while maintaining over 40%
simplification and area-reduction. It consists of one 1:1
peak power-added efficiency (PAE). For example, in [7], transformer, three harmonic tuning capacitors (i.e., 2×Cd
the proposed PA obtains 63% BW1dB, but its peak PAE is

978-1-5386-4545-1/18/$31.00 © 2018 IEEE 220 2018 IEEE Radio Frequency Integrated Circuits Symposium
 Vout+ Vout-
VDD
Cn 1:1 transformer
Vcas CL CL
Lc2

G Cin G Cd Ld CL

M1 M3
S Harmonically S Lc1

D=133μm
Cc Lc1 Vout+
VG Tuned Cd
Vin+
G Rg G Cd W=12μm
Output Ld Ld Vin- Lc2 Vout-
VDD
S
M2 M4 Network S Cc

Ground plane
Cd Ld
G Cin G
Offset=10μm CL

Cn Vcas Vin Vin + -

(a) (b) (c)
Fig. 1. (a) Schematic of the proposed PA. (b) EM model, and (c) simplified schematic of the proposed output network.
input input Lcm1 Lcm2 Z2 input input (1-k2)Lp

Cout Ldm1 Cout Cout Ldm1 Cout

Ld Ld 2Lc1 Ld CL ’
RL k2 Lp RL’
Z1 CL RL 2Lc2 CL
ZL,diff ZL,diff ZL,diff Lp≈Ldm1+Ldm2
Ldm2 ZL,com Ldm2
Cd Cc/2 Cd
Differential-mode Cd ωL≤ω≤ωH
half circuit Common-mode CL’=CL(n/k)2 RL’=RL(k/n)2
half circuit ωL≤ω≤ωH

(a) (b) (c) (d)

input Lk1 1:1/k transformer input input Lcm1 Lcm2 input Lcm1 Lcm2
High impedance
Cout Cout Ldm1 Cout Cout
Lm1 Ld 2Lc1 Ld 2Lc1
Ld
RL CL RL 2Lc2 2Lc2
CL
ZL,diff Ld ZL,diff ZL,com ZL,com
Ldm2 Ldm2 Cd Cc/2 Cd Cc/2
Cd Cd
ω=3ωH Low impedance ω=3ωL
ω=2ωH ω=2ωL
High impedance Low impedance

(e) (f) (g) (h)

Fig. 2. Simplified half circuits of (a) differential-mode, (b) common-mode of the continuous-mode harmonically-tuned output network
at (c) the fundamental frequency and (d) the fundamental equivalent circuit, (e) the 3rd-harmonic of lower band, (f) the 3rd-harmonic of
higher band, (g) the 2nd-harmonic of lower band, and (h) the 2nd-harmonic of higher band.

and Cc), and two matching capacitors (i.e., 2×CL) to realize Fundamental
nd
2 harmonic
rd
3 harmonic
hybrid Class-F and Class-F-1 operations at the lower and 200
higher frequency bands, respectively (Fig. 1b). Note that ωL ωH ωL
3ωH 150 Continuous Continuous
and ωH represent the lowest and highest PA fundamental
Impedance (ohm)

Class-F Class-F-1

frequencies. It also utilizes two symmetrically embedded 100 Ť

2ωL
branches Ld inside the transformer for the 3rd-order 3ωL 50
harmonic impedance tuning in differential-mode, and two
0
extended branches Lc1 and Lc2 for the 2nd-order harmonic 2ωH
impedance tuning in common-mode. The EM model and Load impedance at fundamental
Load impedance at 2nd harmonic
25 30 35 40
Fundametal frequency (GHz)
45

simplified circuit are shown in Fig. 1b and 1c. Compared Load impedance at 3rd harmonic
(a) (b)
with the PA output network in [10], the two matching Fig. 3. (a) Trajectories of half-circuit load impedance at fundamental,
capacitors (i.e., 2×CL) can facilitate the fundamental 2nd- and 3rd-order harmonics (Z0=50Ω), and (b) impedance response.
operation bandwidth extension, and the longer extended
branches Lc1 can provide a larger inductance for 2nd-order TABLE I
harmonic impedance. IMPEDANCE BEHAVIORS FOR CONTINUOUS-MODE
HYBRID CLASS-F AND CLASS-F-1 OPERATIONS
The PA output harmonic termination network is
Frequency Class-F Frequency Class-F-1
explained in Fig. 2(a)-(h). Here, Ldm1/Lcm1 and Ldm2/Lcm2 band band
ZL,F |ZL,F| ZL,FI |ZL,FI|
represent the differential-/common-mode half-circuit
ωL inductive RL,F ωH inductive RL,FI
inductances of the transformer, and the output leads are
2ωL capacitive Low 2ωH capacitive high
absorbed into the secondary coil. The transformer center-
3ωL either high 3ωH either Low
tap is a virtual ground, so Lc1, Lc2 and Cc do not affect the
differential-mode. Cd-Ld-Ldm2 form a multi-resonance tank At fundamental frequencies (i.e., ωL≤ω≤ωH), the series
Z1 (Fig. 2a). In the common-mode half-circuit, the network network Cd-Ld behaves as a small capacitor (Fig. 2c), where
of Cc/2, 2Lc1, and 2Lc2 forms a multi-resonance tank Z2 , as that high impedance branch can be ignored. Thus, ZL,diff can
shown in Fig. 2b. be converted to a simplified model as shown in Fig. 2(d).

221
 20 20 20
Gp DE PAE 60 Gp DE PAE 60 Gp DE PAE 60

50 15 50 50
15 15

DE and PAE (%)

DE and PAE (%)

DE and PAE (%)

Gp=11.4dB 40 40 40
Gp=10.7dB Gp=10.5dB
Gp (dB)

Gp (dB)

Gp (dB)
10 10 30 10 30
30

20
20 20
5 5 5
Psat=18.6dBm 10 10
10 Psat=18.6dBm Psat=18.5dBm
P1dB=16.6dBm P1dB=16.3dBm 0 P1dB=16.3dBm 0
0 0 0 0
-2 0 2 4 6 8 10 12 14 16 18 20 -2 0 2 4 6 8 10 12 14 16 18 20 -2 0 2 4 6 8 10 12 14 16 18 20
CW Output Power Pout (dBm) at 28GHz CW Output Power Pout (dBm) at 37GHz CW Output Power Pout (dBm) at 39GHz
(a) (b) (c)
Fig. 4. Measured CW large-signal performances versus output power P out at (a) 28GHz, (b) 37GHz and (c) 39GHz, respectively.
20 55 20
DC Bias BW1dB=54.3% (23.5-41GHz) S11 S21 S22
50 S21=12.3dB@31GHz
18
45 10
0.55mm 16 BW3dB=51%

S-parameter (dB)
Vin+ Vout+ 40
(25.9-43.7GHz)

PAE (%)
Continuous-
Psat (dBm)

Input 14 PAEpeak=46% 35 0
0.25mm

mode
Matching PA Harmonically-
Network tuned 30
Network 12 Class-F Class-F-1
mode mode 25 -10
Vin- Vout- 10
Psat PAE 20
-20
8 15 20 25 30 35 40 45 50
DC Bias 24 26 28 30 32 34 36 38 40 42
Frequency (GHz)
Frequency (GHz)
Fig. 5. Chip microphotograph. Fig. 6. Measured CW large-signal Fig. 7. Measured small-signal S-
performance versus carrier frequency. parameter versus carrier frequency.

In Fig. 2(d), Lm and Lk are the magnetization and leakage Smith Chart in Fig. 3a. The fundamental load impedance
inductances respectively of the transformer in the half- (ωL≤ω≤ωH) is inductive, while the 2nd harmonic load
circuit differential-mode. The equivalent inductance Lp is impedances (ω=2ωL or ω=2ωH) are capacitive and provide
roughly equal to Ldm1 and Ldm2. Thus, Figure 2d performs -1≤γ<0 for continuous class-F and -1≤ξ<0 for continuous
the four-reactance matching network with the PA output Class-F-1 [11]. Moreover, compared with the fundamental,
capacitance Cout and provides the desired fundamental load the 3rd-oder harmonic load impedance is low for lower band
impedance (i.e., ωL≤ω≤ωH) to the PA. (ω=3ωL) while it is high for higher band (ω=3ωH). The load
At the 3rd-order harmonic of the higher band (i.e., impedance trajectories demonstrate the continuous-mode
ω=3ωH), the series network Cd-Ld impedance is slightly hybrid Class-F/F-1 operations. Also, the impedance of each
below its series resonance, which shorts out Ldm2 and forms harmonic is presented in Fig. 3b. The continuous-mode
a series resonance of Cd-Ld-Lm1-Lk1 to produce a low load operations for each harmonic are shown in Table I [11].
impedance (Fig. 2e). In this case, Lm1 and Lk1 represent the
magnetization and leakage inductances of coil Ldm1 of the III. MEASUREMENT RESULTS
transformer in the half-circuit differential-mode. Also, at
the 3rd harmonic of the lower band (i.e., ω=3ωL), ZL,diff sees Our PA design is implemented in a standard 45nm
a high impedance by Ldm1 and Z1 in parallel with Cout. CMOS SOI process with a 0.55×0.25mm2 core area as
At the 2nd-order harmonic of the higher band (i.e., shown in Fig. 5. The PA chip is measured by direct probing.
ω=2ωH), Z2 provides a high impedance, while the Figure 4 shows the measured CW large-signal
remaining Cd-Ld series tank behaves as a capacitor (Fig. 2g). performance at 28, 37 and 39GHz respectively. At 28GHz,
Therefore, with proper tuning, the 2 nd-order harmonic the proposed PA achieves 18.6dBm saturated power Psat,
impedance ZL,com is dominated by Cout, Lcm1 and the 45.7% peak PAE, and 11.4dB power gain Gp. At 37GHz,
effective capacitance due to series Cd-Ld, achieving the the PA demonstrates 18.6dBm Psat, 40.2% peak PAE, and
desired continuous-mode 2nd-order harmonic impedance. 10.7dB Gp. At 39GHz, the PA achieves 18.5dBm Psat,
Additionally, at the 2nd-order harmonic of the lower band 41.2% peak PAE, and 10.5dB Gp. The measured PAE
(i.e., ω=2ωL), Z2 becomes inductive. Moreover, the series includes the PA stage and the loss of the output network.
network Cd-Ld remains capacitive. Therefore, ZL,com can Our PA achieves very high efficiency (i.e., peak PAE≥40%)
present a low overall impedance with proper tuning. and delivers almost constant Psat (i.e., Psat variation is within
The trajectories of the half-circuit load impedance at 0.1dB) at all the potential mm-Wave 5G bands [1].
fundamental, 2nd- and 3rd-order harmonics with the The measured CW large-signal performance versus
absorbed PA output capacitance Cout are shown on the frequency is shown in Fig. 6. The Psat output power 1dB
bandwidth (BW1dB) is 54.3% from 23.5GHz to 41GHz. The

222
 TABLE II
COMPARISON OF STATE-OF-THE-ART SILICON-BASED MM-WAVE PAS AT RELATED FREQUENCIES
Pout 1dB Pout 1dB Core size
Author Psat (dBm) PAE (%) Gain (dB) P1dB (dBm) Process Topology
Frequency (GHz) Bandwidth (%) (mm2)
18.6@28GHz 45.7@28GHz 11.4@28GHz 16.6@28GHz
45nm SOI Hybrid of Continuous
This work 23.5-41 54.3 18.6@37GHz 40.2@37GHz 10.7@37GHz 16.3@37GHz 0.14
CMOS Class-F/F-1
18.5@39GHz 41.2@39GHz 10.5@39GHz 16.3@39GHz
RFIC 2017
26-34 26.7 14.75 46.4 10 13.2 0.12 65nm CMOS Continuous Class-F
S. Ali
ISSCC 2018
19-29.5 43.3 17 43.5 20 15.2 0.29 130nm SiGe Continuous Class-F-1
T. Li
T-MTT 2017
27-39 7.1 18.8 35.3 15.5 15.9 0.27 130nm SiGe Continuous Class-AB
A. Sarkar

RFIC 2017 Class-AB w/ Output

25-48* 63* 16.6 24.2 20.8 13.4 0.16 28nm CMOS
M. Vigilante Power Combiner

RFIC 2017
28-33* 22.2* 19.8 21 22 16 0.59 28nm CMOS Doherty
P. Indirayanti
16.8@28GHz 20.3@28GHz 18.2@28GHz 15.2@28GHz
ISSCC 2017
28-42 40 17.1@37GHz 22.6@37GHz 17.1@37GHz 15.5@37GHz 1.76 130nm SiGe Doherty
S. Hu
17@39GHz 21.4@39GHz 16.6@39GHz 5.4@39GHz
ISSCC 2014
24-31 25.5 17.1 40 10.3 15 0.27 130nm SiGe Hybrid of Class-F-1/F
S. Mortazavi
*Graphically estimated from reported figures

peak PAE is 46% at 29GHz. Also, it maintains over 30% Class-F/F-1 power amplifier in CMOS. Moreover, this
PAE from 25.5GHz to 41GHz (BW=46.6%), which design is the first PA to maintain more than 40% PAE, and
addresses the efficiency-bandwidth challenge of mm-Wave consistently deliver more than 18dBm to cover multiple
PAs. In addition, Figure 6 also demonstrates that the Class- mm-Wave 5G systems [1].
F mode operates around 28GHz, while Class-F-1 mode
operates around 38GHz. A mode transition is clearly shown ACKNOWLEDGEMENT
at 35GHz, which matches our analysis in Section II. Figure
The authors thank GlobalFoundries for chip fabrication.
7 depicts the measured small-signal S-parameter versus
frequency. It shows that the small-signal 3dB bandwidth REFERENCES
(BW3dB) is 51% from 25.9GHz to 43.7GHz. Our design also
exhibits good input matching, achieving return loss (RL) [1] Huawei, “5G Spectrum Public Policy Position,”
http://www.huawei.com/en/about-huawei/public-
below 10dB from 27.3 to 43GHz. The peak S21 is 12.3dB at policy/5g-spectrum.
31GHz. The performance comparison with other silicon- [2] A. Sarkar, et al., “A 28-GHz Harmonic-Tuned Power
based PAs at the related frequencies is shown in Table II. Amplifier in 130-nm SiGe BiCMOS,” T-MTT, 2017.
[3] S. Ali, et al., “A 42–46.4% PAE continuous class-F power
IV. CONCLUSION amplifier with Cgd neutralization at 26–34 GHz in 65 nm
CMOS for 5G applications,” RFIC, 2017.
In this paper, we propose a differential fully-integrated [4] S. Mortazavi, et al., “Integrated Inverse Class-F Silicon
continuous-mode hybrid Class-F/F-1 power amplifier, Power Amplifiers for High Power Efficiency at Microwave
implemented in a standard 45nm CMOS SOI. This PA and mm-Wave,” JSSC, 2016.
utilizes our proposed continuous-mode harmonically tuned [5] S. Mortazavi, et al., “A class F-1/F 24-to-31GHz power
output network to provide the required 2nd- and 3rd-order amplifier with 40.7% peak PAE, 15dBm OP1dB, and
50mW Psat in 0.13μm SiGe BiCMOS,” ISSCC, 2014.
harmonic impedance terminations for continuous Class-F [6] S. Hu, et al., “2.1 A 28GHz/37GHz/39GHz multiband
and continuous Class-F-1 operations respectively. Our linear Doherty power amplifier for 5G massive MIMO
design achieves peak PAE of 46% and Psat output power applications,” ISSCC, 2017.
1dB bandwidth of 54.3% from 23.5GHz to 41GHz. In [7] M. Vigilante, et al., “A 29-to-57GHz AM-PM
addition, regarding the potential 5G bands, e.g., 28GHz, compensated class-AB power amplifier for 5G phased
arrays in 0.9V 28nm bulk CMOS,” RFIC, 2017.
37GHz and 39GHz [1], the peak PAE values are all higher [8] Y. Zhang, et al., ”A high-efficiency linear power amplifier
than 40% and the Psat values are all greater than 18.5dBm. for 28GHz mobile communications in 40nm
Also, our continuous-mode harmonically-tuned output CMOS,” RFIC, 2017.
network only occupies one single transformer footprint and [9] B. Park, et al., “Highly Linear mm-Wave CMOS Power
does not require any additional tunable elements or Amplifier,” T-MTT, 2016.
[10] T. Li, et al., “A continuous-mode harmonically tuned 19-
switches, providing an ultra-compact design, i.e., the chip to-29.5GHz ultra-linear PA supporting 18Gb/s at 18.4%
core size is only 0.14mm2 for a differential design. modulation PAE and 43.5% peak PAE,” ISSCC, 2018.
To the best of our knowledge, this design is the first [11] K. Chen, et al., “Design of Broadband Highly Efficient
demonstration of a fully integrated continuous-mode hybrid Harmonic-Tuned Power Amplifier Using In-Band
Continuous Class-F/F-1 Mode Transferring,” T-MTT, 2012.

223