electronics

Article
Effective Digital Predistortion (DPD) on a Broadband
Millimeter-Wave GaN Power Amplifier Using LTE
64-QAM Waveforms
Gokul Somasundaram 1, *, Jill C. Mayeda 1 , Clint Sweeney 1 , Donald Y. C. Lie 1 and Jerry Lopez 1,2

1 Department of Electrical and Computer Engineering, Texas Tech University, Lubbock, TX 79409, USA;
[email protected] (J.C.M.); [email protected] (C.S.); [email protected] (D.Y.C.L.);
[email protected] (J.L.)
2 Noise Figure Research, Renton, WA 98057, USA
* Correspondence: [email protected]

Abstract: We demonstrate in this work effective linearization on a millimeter-wave (mm-Wave)

broadband monolithic gallium nitride (GaN) power amplifier (PA) using digital predistortion (DPD).
The PA used is a two-stage common-source (CS)/2-stack PA that operates in the mm-Wave 5G
FR2 band, and it is linearized with the generalized memory polynomial (GMP) DPD and tested
using 4G (4th generation) long-term-evolution (LTE) 64-QAM (quadrature amplitude modulation)
modulated signals with a PAPR (peak-to-average power ratio) of 8 dB. Measurement results after
implementing GMP DPD indicate considerable broadband improvement in the adjacent channel
leakage power ratio (ACLR) of 16.9 dB/17.3 dB/16.5 dB/15.1 dB at 24 GHz/28 GHz/37 GHz/39 GHz,
respectively, with a common average POUT of 15 dBm using a 100 MHz LTE 64-QAM input signal. At
a fixed frequency of 28 GHz, the GaN PA after GMP DPD achieved signal bandwidth-dependent
ACLR improvement and root-mean-square (rms) EVM (error vector magnitude) reduction using
20 MHz/40 MHz/80 MHz/100 MHz LTE 64-QAM waveforms with a common average POUT of
15 dBm. The GaN PA thus achieved very good linearization results compared to that in other state-
of-the-art mm-Wave PA DPD studies in the literature, suggesting that GMP DPD should be rather
Citation: Somasundaram, G.; effective for linearizing mm-Wave 5G broadband GaN PAs to improve POUT, Linear .
Mayeda, J.C.; Sweeney, C.; Lie,
D.Y.C.; Lopez, J. Effective Digital
Keywords: 4G (4th generation); 5G (5th generation); adjacent channel leakage power ratio (ACLR);
Predistortion (DPD) on a Broadband
carrier aggregation (CA); digital predistortion (DPD); error vector magnitude (EVM); gallium nitride
Millimeter-Wave GaN Power
(GaN); generalized memory polynomial (GMP); long-term evolution (LTE); millimeter-wave (mm-
Amplifier Using LTE 64-QAM
Wave); power amplifier (PA); quadrature amplitude modulation (qam)
Waveforms. Electronics 2023, 12, 2869.
https://doi.org/10.3390/
electronics12132869

Academic Editor: Reza K. Amineh 1. Introduction

Received: 17 May 2023 The 5th generation (5G) mobile network currently offers both the sub-6 GHz frequency
Revised: 19 June 2023 range 1 (FR1) band and the millimeter-wave (mm-Wave) frequency Range 2 (FR2) band of
Accepted: 27 June 2023 24.25–52.6 GHz to achieve up to 10 Gb/s download speeds for enhanced mobile broadband
Published: 28 June 2023 (eMBB) and other applications [1]. For mm-Wave power amplifiers (PA) designed at
the higher 5G FR2 band, the trade-off in linearity vs. power-added efficiency (PAE) can
become considerably more complicated, making it more difficult and increasingly expensive
to construct mm-Wave 5G small cells and base stations (i.e., macrocells, microcells and
Copyright: © 2023 by the authors.
picocells) [2,3]. For a mm-Wave RF transmitter (TX), the performance of the PA can
Licensee MDPI, Basel, Switzerland.
often dominate the overall TX performance in heat dissipation, POUT , linearity, efficiency,
This article is an open access article
distributed under the terms and
reliability, etc. For example, modulated input signals with high a peak-to-average power
conditions of the Creative Commons
ratio (PAPR) can degrade a PA’s average PAE and linear output power, POUT, Linear , and
Attribution (CC BY) license (https:// cause TX overheating [4]. Therefore, a highly efficient PA design with high POUT, Linear is
creativecommons.org/licenses/by/ crucial for 5G applications. Digital predistortion (DPD) can play a pivotal role here as
4.0/). it can linearize the mm-Wave PAs and significantly improve their POUT, Linear and design

Electronics 2023, 12, 2869. https://doi.org/10.3390/electronics12132869 https://www.mdpi.com/journal/electronics

Electronics 2023, 12, 2869 2 of 19

margins. The effectiveness of various DPD techniques can be evaluated based on their
improvement of the adjacent channel leakage power ratio (ACLR) in the PA’s output
spectrum, and the modulation quality determined from constellation diagrams (i.e., error
vector magnitude (EVM)) [2]. Many DPD techniques have been reported in recent years,
such as the popular look-up table (LUT)-based DPD [5], memory polynomial DPD [6],
generalized memory polynomial DPD [7] and high I/Q imbalance DPD [8]. For example,
the DPD model reported in [8] relies on constructing the multi-toned signal consisting of the
I (in-phase) signal whose frequency components do not correlate with the Q (quadrature)
signal components. Using interleaved multi-toned signals allows a faster detection of the
predistortion coefficients, and the coefficients of the I/Q model are estimated and compared
until the error is zero [8]. Another possible benefit of using multi-tone signals is that the
output signals can have a high spurious free dynamic range (SFDR).
This work, however, uses a relatively simpler single-tone GMP DPD model imple-
mented in a user-friendly National Instruments (NI) mm-Wave PA testbench and not
needing external FPGAs (field programmable gate arrays) [5]. This work applies a GMP
DPD algorithm to linearize a two-stage common source (CS)/two-stack gallium nitride
(GaN) monolithic mm-Wave PA designed in our lab [9]. This GMP DPD achieves im-
pressive performance on the broadband mm-Wave PA, compared to that of other similar
mm-Wave PA studies in the literature [6–8], as discussed in Section 2. The DPD algorithm
and the experimental mm-Wave hardware DPD testbench setup are also discussed in
Section 2. Section 3 presents the measurement results of the PA tested using long-term Evo-
lution (LTE) quadrature amplitude modulation (64-QAM) modulated signals of different
signal bandwidths (BW) at various output power levels and frequencies. We conclude by
comparing our work with that conducted on several state-of-the-art mm-Wave DPD PAs in
the literature.

2. Proposed GMP DPD Algorithm and Mm-Wave Hardware Testbench Setup

The memoryless look-up table (LUT) is an effective DPD technique with which to
linearize mm-Wave PAs, but it does not consider the memory effects of the PAs, which can
cause significant nonlinearity, such as that reported for many GaN PAs [10–13]. Therefore,
generalized memory polynomial (GMP) DPD is used in this work to linearize our mm-
Wave GaN PA [9]. In this case, the output power of the PA is not only a function of the
present input signal, but is also dependent on the past input signal. The output z(n) from
pre-distorter model (A) is a non-linear function of input x(n − q) for all q > 0, and it can be
written as Equation (1) [2]:

z ( n ) = h (0 ) + ∑ q1 h 1 ( q 1 ) × x ( n − q 1 )
+ ∑ q1 ∑ q2 h 2 ( q 1 , q 2 ) × x ( n − q 1 ) × x ( n − q 2 ) (1)
+···

Here, h(q) represents voltage kernels, q is the memory depth, and x(n) is the input
signal to the PA. h(0) is the system response for x(n) = 0, h1 (q1 ) is the unit impulse response
of the system and h2 (q1 , q2 ) is the impulse response of the quadratic component of the
system [14]. h(0) can be eliminated from Equation (1) as this is a memory polynomial. The
z(n) is a non-linear function for all (all past inputs), and n is number of iterations. The term
x(n) can be represented in the form of a complex baseband as in Equation (2) [15]:
 
x (n) = e jnωo × Re{ x (n)} + e− jnωo × Im{ x (n)} /2 (2)

Here, ωo = 2π f o and f o is the center frequency; in most cases, we consider compo-

nents that are close to f o . The past input signals to the PA are included in terms of voltage
kernels along with the present input signal, x(n), to the PA. In this manner, Equation (1)
digitally compensates for PA’s memory effects and nonlinearity. The GMP algorithm is a
common model from the modified Volterra series using Equation (1) [2]. The GMP model
is more sophisticated than the LUT DPD model [5,16,17] is in terms of its configurations
 𝑧𝑧(𝑛𝑛) = ∑𝑘𝑘=0 ∑𝑞𝑞=0 𝑎𝑎𝑘𝑘,𝑞𝑞 × 𝑥𝑥[𝑛𝑛 − 𝑞𝑞 ] × |𝑥𝑥[𝑛𝑛 − 𝑞𝑞 ]| +
∑𝑘𝑘=1 ∑𝑄𝑄𝑞𝑞=0
𝐾𝐾𝑏𝑏 𝑏𝑏
∑𝑀𝑀 𝑘𝑘
𝑀𝑀=1 𝑏𝑏𝑘𝑘,𝑞𝑞,𝑀𝑀 × 𝑥𝑥 [𝑛𝑛 − 𝑞𝑞 ] × |𝑥𝑥 [𝑛𝑛 − 𝑞𝑞 − 𝑀𝑀]| +
𝑏𝑏
(8)
∑𝐾𝐾 𝑄𝑄𝑐𝑐 𝑀𝑀𝑐𝑐
𝑘𝑘=1 ∑𝑞𝑞=0 ∑𝑀𝑀=1 𝑐𝑐𝑘𝑘,𝑞𝑞,𝑀𝑀 × 𝑥𝑥 [𝑛𝑛 − 𝑞𝑞 ] × |𝑥𝑥 [𝑛𝑛 − 𝑞𝑞 + 𝑀𝑀]|
𝑐𝑐 𝑘𝑘

Here,
Electronics 2023, 12, 2869 𝐾𝐾𝑎𝑎 , 𝑄𝑄𝑎𝑎 are the number of coefficients of a generalized form of an aligned enve- 3 of 19
lope in the memory polynomial function, 𝐾𝐾𝑏𝑏 , 𝑄𝑄𝑏𝑏 , 𝑀𝑀𝑏𝑏 are the number of coefficients of the
lagging envelope in the memory polynomial function, and 𝐾𝐾𝑐𝑐 , 𝑄𝑄𝑐𝑐 , 𝑀𝑀𝑐𝑐 are the number of
coefficients of the leading
which comeenvelope in theorder
with a higher memory polynomial
and address memory function.
effects.This
From makes
Figurethe 1, we can see
coefficients simple thatand
the easy
inputtosignal,
estimate
x(n),from any
is sent least-squares
into algorithm.
the pre-distorter model In (A)most cases,the voltage
to include
Equation (8) [2,15] hasfunction,
kernel a higher domination
h(q), of odd-order
which accounts linearities
for the memory effects.than
The of even-order
output signal from PA y(n)
linearities. As the DPD models are interested in terms that are closer to frequencyẑ(n)
goes through the 1/G block to the pre-distorter (A2) which is ẑ(n). This 𝑓𝑓𝑜𝑜 can
to be written
as y(n)/G. The error signal, e(n), is estimated per Equation (3) [15]. Both the pre-distorters
yield satisfactory performance, the odd order linearities are chosen [15] and it may not be
(A and A2) have the same DPD model, Equation (1) [15].
necessary to implement all the delay terms in it [15]. Here, 𝐾𝐾𝑎𝑎 , 𝑄𝑄𝑎𝑎 ; 𝐾𝐾𝑏𝑏 , 𝑄𝑄𝑏𝑏 , 𝑀𝑀𝑏𝑏 ; 𝐾𝐾𝑐𝑐 , 𝑄𝑄𝑐𝑐 , 𝑀𝑀𝑐𝑐
are the index arrays of the aligned, lagging, andeleading (n) = envelopes
y(n) in the memory poly-
G − z(n)
nomial function of Equation (8) [2,15]. = ẑ(n) − z(n)
(3)

x(n) z(n) y(n)

Pre-distorter(A) Power Amplifier
e(n) 1/G
−+
(n)
Pre-distorter(A2)
Copy of Pre-distorter (A)

Figure 1. Block diagram

Figure 1.ofBlock
the GMP DPD
diagram of model
the GMP with
DPDvoltage kernel
model with functions
voltage kernelused for this
functions usedwork
for this work [2].
[2].
The pre-distorter model (A2) has y(n)/G as its input, where G is the PA’s gain. The
Figure 2 represents
output ofthe themm-Wave
A2 model PA testing
is ẑ(n). Thesetup in our lab
pre-distorter used(A2)
model to linearize
is a copythe PA actual pre-
of the
with DPD, which consists
distorter (A)ofwhich
the I/Qhassignal
x(n) asgenerator PXIe-5820
its input and z(n) as its[19] (PXIeThe
output. stands for the
algorithm renders both
PCI express extensions
the output forz(n)
instrumentation) and the
and ẑ(n) of pre-distorter up-converter
models (A and A2)PXIe-3622 by NI
until the error (la-e(n), is zero.
signal,
beled “B” in Figure 2). Baseband
As the user defines 64-QAM
x(n), theLTEdatainput signals
samples of four by
are known different BWs (1 The
the algorithm. × 20input of the
MHz/2 × 20 MHz/4 × 20 MHz/5 × 20 MHz) with PAPR = 8 dB and a sampling rate of x(n). Once
pre-distorter model (A2) is y(n)/G, which is almost like the initial input signal,
the signals are sent into the pre-distorter models (A and A2), data samples of both ẑ(n) and
around 400 MHz–500 MHz are generated using the NI-RFmx waveform creator (labeled
z(n) are gathered with the help of the NI-RFmx GUI (graphical user interface), which will
“A” in Figure 2). Input signals of BW of 40 MHz/80 MHz/100 MHz are created by carrier
be discussed later [18]. The pre-distorter coefficients are found for both outputs of models
aggregation (CA) of the baseband 64-QAM 1 × 20 MHz input signal using the NI-RFmx
(A and A2) and compared until no changes are detected (i.e., e(n) is zero). If the NI-RFmx
waveform creator. GUIFor example,
detects to create
a difference in the 100 MHz
coefficients signal,
(i.e., e(n) 6=the0)number of carrier
, the process offset
continues by updating
signals is increased from 1 (1 × 20 MHz) to 5 (5 × 20 MHz). NI-RFmx
the polynomial order (k) of the DPD models until the coefficients become is a software GUI
the same (i.e.,
that optimizes NI-RF
e(n) = instrumentations
0). In Equation (1),using h(0) isvarious standard
eliminated as it issignals.
the system NI-RFmx
response contains
of x(n − q) ∀ q = 0.
a pile of software programs
From Equationthat (2), can
z(n) be
canused to generate
be written in termsmodulated signal (envelope)
of the polynomial waveformsorder “k” of
using the NI-RFmx waveform
the PA as showncreator and perform
in Equation (4) [15]: DPD using the NI-RFmx panel [2,19].
The generated baseband signal waveform is then sent into I/Q signal generator NI PXIe-
5820. The baseband signal is n) = ∑qQ1 =into
z(converted 0 h1 (an
q1 )intermediate Q
∑q1 = 0,q2 =(IF)
× x (n − q1 ) +frequency 0 h2which
(q1 , q2 ) is× then
x (n
Q
sent to the up converter NI−PXIe-3622
q1 ) × x (n −toq2be
) +converted
· · · + ∑q1 =into
0,q2 the desired
= 0,...,q k = 0
hRF/mm-Wave
k ( q1 , q2 , . . . , qk ) signal
× x (n− (4)
frequency. With the help of mm radio head 5582 q1 )× x(labeled
(n − q2 ) “C”
. . . x (in
n−Figure
qk ) 2), the intended
mm-Wave frequency is passed into a driver amplifier (labeled “D” in Figure 2) to drive
If q = 0, Equation (4) turns out to be a memoryless polynomial. To consider it as a
the two-stage CS/2-stack GaN PA (labeled “E” in Figure 2) [9] at the desired input signal.
memory polynomial equation, here, hk (q1 , q2 , ..qk )= 0 ∀q1 6= q2 , q2 6= q3 and qk−1 6= qk ,
The broadband driver amplifier used has a BW of 18–47 GHz and it is from RF-Lambda
which can be simplified into Equation (5):
(RFLUPA18G47GC PA with G = 37 dB and POUT, MAX = 31 dBm) [20]. The amplified RF/mm-
z(n) = ∑qQ= 0 (h1 (q) × x (n − q) + h2 (q, q) × x (n − q) × | x (n − q)|+
(5)
· · · + hk (q, q, . . . , q) × x (n − q) × | x (n − q)|k−1 )

From Equation (5) [15,19], the co-efficient from the simplified Volterra series is written
as follows:
K −1
∑k = 0 ak,q = h1 (q) + h2 (q, q) + · · · + hk (q, q, . . . , q) (6)
Electronics 2023, 12, 2869 4 of 19

The final simplified form of the memory polynomial model equation used in the
pre-distorter for the Kth order is thus expressed in Equation (7) below:
K −1 Q
z(n) = ∑k = 0 ∑q = 0 ak,q × x[n − q] × |x[n − q]|k (7)

Considering the lead and lag memory polynomial (envelope) orders in the memory
polynomial function, multiple delayed versions of both positive and negative cross-term
time shifts result in GMP DPD model Equation (7) [15]. In most cases, the memory depth (q)
from Equation (7) can be set by the user. For the generalized memory polynomial function,
lead and lag memory depths are defined below:

z(n) = ∑kKa=−01 ∑qQ=

a k
0 ak,q × x [ n − q ] ×| x [ n − q ]| +
K Q M k
∑k b= 1 ∑q b= 0 ∑ Mb= 1 bk,q,M × x [n − q] × | x [n − q − M ]| + (8)
K Q M k
∑k c= 1 ∑q c= 0 ∑ Mc= 1 ck,q,M × x [n − q] ×| x [n − q + M]|

Here, Ka , Q a are the number of coefficients of a generalized form of an aligned enve-

lope in the memory polynomial function, Kb , Qb , Mb are the number of coefficients of the
lagging envelope in the memory polynomial function, and Kc , Qc , Mc are the number of
coefficients of the leading envelope in the memory polynomial function. This makes the
coefficients simple and easy to estimate from any least-squares algorithm. In most cases,
Equation (8) [2,15] has a higher domination of odd-order linearities than of even-order
linearities. As the DPD models are interested in terms that are closer to frequency f o to
yield satisfactory performance, the odd order linearities are chosen [15] and it may not be
necessary to implement all the delay terms in it [15]. Here, Ka , Q a ; Kb , Qb , Mb ; Kc , Qc , Mc are
the index arrays of the aligned, lagging, and leading envelopes in the memory polynomial
function of Equation (8) [2,15].
Figure 2 represents the mm-Wave PA testing setup in our lab used to linearize the
PA with DPD, which consists of the I/Q signal generator PXIe-5820 [19] (PXIe stands
for the PCI express extensions for instrumentation) and the up-converter PXIe-3622 by
NI (labeled “B” in Figure 2). Baseband 64-QAM LTE input signals of four different BWs
(1 × 20 MHz/2 × 20 MHz/4 × 20 MHz/5 × 20 MHz) with PAPR = 8 dB and a sampling
rate of around 400 MHz–500 MHz are generated using the NI-RFmx waveform creator
(labeled “A” in Figure 2). Input signals of BW of 40 MHz/80 MHz/100 MHz are created
by carrier aggregation (CA) of the baseband 64-QAM 1 × 20 MHz input signal using the
NI-RFmx waveform creator. For example, to create the 100 MHz signal, the number of
carrier offset signals is increased from 1 (1 × 20 MHz) to 5 (5 × 20 MHz). NI-RFmx is
a software GUI that optimizes NI-RF instrumentations using various standard signals.
NI-RFmx contains a pile of software programs that can be used to generate modulated
signal waveforms using the NI-RFmx waveform creator and perform DPD using the NI-
RFmx panel [2,19]. The generated baseband signal waveform is then sent into I/Q signal
generator NI PXIe-5820. The baseband signal is converted into an intermediate frequency
(IF) which is then sent to the up converter NI PXIe-3622 to be converted into the desired
RF/mm-Wave signal frequency. With the help of mm radio head 5582 (labeled “C” in
Figure 2), the intended mm-Wave frequency is passed into a driver amplifier (labeled “D”
in Figure 2) to drive the two-stage CS/2-stack GaN PA (labeled “E” in Figure 2) [9] at the
desired input signal. The broadband driver amplifier used has a BW of 18–47 GHz and it
is from RF-Lambda (RFLUPA18G47GC PA with G = 37 dB and POUT, MAX = 31 dBm) [20].
The amplified RF/mm-Wave signal from the PA is sent back to the down converter NI
PXIe-3622. The IF signals are sent back to PXIe-5820, which are passed into the lab’s PC
(personal computer) for linearization. The polynomial coefficients ak,q , bk,q,M , ck,q,M are
estimated from the received signal into the PC, and are used to linearize/eliminate the
difference vs. the source baseband signal. Settings of both the order and depth of terms
such as the polynomial order, memory lag, and memory lead can be adjusted in the NI-
RFmx panel [2]. For this linearization experiment, a polynomial order (k) of 7, memory lag
 Wave signal from the PA is sent back to the down converter NI PXIe-3622. The IF signals
are sent back to PXIe-5820, which are passed into the lab’s PC (personal computer) for
linearization. The polynomial coefficients 𝒂𝒂𝒌𝒌,𝒒𝒒 , 𝒃𝒃𝒌𝒌,𝒒𝒒,𝑴𝑴 , 𝒄𝒄𝒌𝒌,𝒒𝒒,𝑴𝑴 are estimated from the re-
ceived signal into the PC, and are used to linearize/eliminate the difference vs. the source
Electronics 2023, 12, 2869
baseband signal. Settings of both the order and depth of terms such as the polynomial 5 of 19

order, memory lag, and memory lead can be adjusted in the NI-RFmx panel [2]. For this
linearization experiment, a polynomial order (k) of 7, memory lag (𝑸𝑸𝒃𝒃 ) and lead depth
((𝑸𝑸 of 5lead
Qb𝒄𝒄)) and anddepth
memory of 5(𝑴𝑴
(Q c )lag and and leadlagorder
𝒃𝒃 ) memory ( Mb )(𝑴𝑴
and of 2order
𝒄𝒄 )lead are used.
(M c ) A
of 2lead and lag
are used. A
memory
lead depth
and lag of “2” depth
memory is commonly
of “2” isused for the GMP
commonly DPD
used for theexperiments in the literature
GMP DPD experiments in
[7,8].
the In our study,
literature however,
[7,8]. In our study, a lead
however,and lag memory
a lead and lagdepth of “5”depth
memory was used,
of “5”which may
was used,
have made our GMP DPD have better performance than those in the literature while tack-
which may have made our GMP DPD have better performance than those in the literature
ling wideband
while signals. Itsignals.
tackling wideband took less thanless
It took 30 sthan
to perform the GMP
30 s to perform theDPD
GMPwithDPDthe higher
with the
depths in our work. The GMP DPD used in this paper was also applied in a very user-
higher depths in our work. The GMP DPD used in this paper was also applied in a very
friendly NI testbench.
user-friendly NI testbench.

A C
(i) (ii)
B

E
D

Figure2.2. Mm-wave
Figure Mm-wave PAPA DPD
DPD experimental
experimentalsetup
setupused
usedininour
ourlab:
lab: (A)
(A) PC;
PC; (B)
(B) NI
NI PXIe;
PXIe; (i)
(i) PXIe-5820;
PXIe-5820;
(ii) PXIe-3622; (C) mm-Wave radio head 5582; (D) Driver; (E) PA.
(ii) PXIe-3622; (C) mm-Wave radio head 5582; (D) Driver; (E) PA.

Whenthe
When theuserusercreates
createsaabaseband
basebandwaveform
waveformwith withthe
thehelp
helpofofthe
the NI-RFmx
NI-RFmxwaveform waveform
creatorbased
creator based on on thethe specification,
specification, the the NI-RFmx
NI-RFmx panel
panel saves
saves the
the reference
reference waveform
waveform con- con-
figurations to compare with the received baseband waveform
figurations to compare with the received baseband waveform after amplification. Both after amplification. Both the
waveforms
the waveforms areare loaded
loaded intointo
LabVIEW
LabVIEW using NI-RFmx.
using NI-RFmx.Then, both
Then, waveforms
both waveforms undergo undergo the
mathematical
the mathematical Equation
Equation (8) (8)
[15][15]
to derive thethe
to derive polynomial
polynomialcoefficients
coefficients𝒂𝒂𝒌𝒌,𝒒𝒒
ak,q , 𝒃𝒃,𝒌𝒌,𝒒𝒒,𝑴𝑴
bk,q,M , 𝒄𝒄𝒌𝒌,𝒒𝒒,𝑴𝑴
, ck,q,M as
allall
as otherotherfactors factorsare areeither
eithergiven
givenbyby the
the user
user or derived from
or derived from the
the waveforms.
waveforms. Here, Here,
a𝒂𝒂k,q
𝒌𝒌,𝒒𝒒, ,b𝒃𝒃k,q,M
𝒌𝒌,𝒒𝒒,𝑴𝑴,, c𝒄𝒄k,q,M is aa transpose
𝒌𝒌,𝒒𝒒,𝑴𝑴 is transposecolumn
columnvector
vectormatrix
matrixasasshown
shownin inEquation
Equation(9) (9)[15]:
[15]:

𝑎𝑎𝑘𝑘,𝑞𝑞 = (𝑎𝑎𝑞𝑞0, 𝑎𝑎𝑞𝑞1 , . . . , 𝑎𝑎𝑞𝑞(𝑘𝑘𝑎𝑎−1))T𝑇𝑇

ak,q = aq0 , aq1 , ..., aq(k a −1)
 T 𝑇𝑇
𝑏𝑏𝑘𝑘,𝑞𝑞,𝑀𝑀b = (𝑏𝑏=𝑚𝑚𝑚𝑚0

, 𝑏𝑏 , b , ., ....,
bmq0𝑚𝑚𝑚𝑚1
. , b𝑏𝑏𝑚𝑚𝑚𝑚(𝑘𝑘𝑏𝑏−1) ) (9)
(9)
k,q,M mq1 mq(k b −1)
 T 𝑇𝑇
𝑐𝑐𝑘𝑘,𝑞𝑞,𝑀𝑀c = (𝑐𝑐 𝑚𝑚𝑚𝑚0 , 𝑐𝑐𝑚𝑚𝑚𝑚1 , . . . , 𝑐𝑐𝑚𝑚𝑚𝑚(𝑘𝑘 −1) )

k,q,M = cmq0 , cmq1 , ..., cmq(k c −𝑐𝑐1)
The transposition of the column vector matrix for the polynomial coefficients is car-
ried The
out astransposition of perform
it is easier to the column vector
vector matrix for the
multiplication polynomial
with the x(n −coefficients is carried
q). Considering only
out as it is easier to perform vector multiplication with the x(n − q).
the higher polynomial order is sent to LabVIEW, Equation (9) is as follows: Considering only the
higher polynomial order is sent to LabVIEW, Equation (9) is as follows:
 T
ak,q = a q ( k a −1)
 T
bk,q,M = bmq(kb −1) (10)
 T
ck,q,M = cmq(kc −1)
 𝑎𝑎𝑘𝑘,𝑞𝑞 = (𝑎𝑎𝑞𝑞(𝑘𝑘𝑎𝑎−1) )𝑇𝑇
𝑏𝑏𝑘𝑘,𝑞𝑞,𝑀𝑀 =(𝑏𝑏𝑚𝑚𝑚𝑚(𝑘𝑘𝑏𝑏−1) )𝑇𝑇 (10)
Electronics 2023, 12, 2869
𝑐𝑐𝑘𝑘,𝑞𝑞,𝑀𝑀 = (𝑐𝑐𝑚𝑚𝑚𝑚(𝑘𝑘𝑐𝑐−1) )𝑇𝑇 6 of 19

Figure 3 shows the palette of the GMP DPD operation used in the NI’s LabVIEW
environment, where the orders
Figure are given
3 shows by the
the palette of user and the
the GMP DPD reference
operationbaseband wave-
used in the NI’s LabVIEW
form is loaded from the NI-RFmx.
environment, where The
theuser is restricted
orders are given by from
theaccessing
user and the the reference
palette asbaseband
the wave-
operation happens under
form the hood
is loaded fromofthetheNI-RFmx.
NI-RFmx The GUI.userInitially, only the
is restricted from reference base-
accessing the palette as
band waveform is theloaded intohappens
operation the highlighted
under theGMP hoodDPDof thepalette
NI-RFmx to obtain the reference
GUI. Initially, only the reference
polynomial coefficients, which are stored as the output error signals. Once we obtain thethe reference
baseband waveform is loaded into the highlighted GMP DPD palette to obtain
amplified signal polynomial
back to NI PXIecoefficients, which are stored as the output error signals. Once we obtain the
as shown in Figure 2, it is then sent into the “Instrument
amplified signal back to NI PXIe as shown in Figure 2, it is then sent into the “Instrument
Handle In” shown in Figure 3 as the received baseband waveform and to the highlighted
Handle In” shown in Figure 3 as the received baseband waveform and to the highlighted
GMP DPD palette to derive the polynomial coefficients 𝑎𝑎 coefficients
GMP DPD palette to derive the polynomial𝑘𝑘,𝑞𝑞
, 𝑏𝑏𝑘𝑘,𝑞𝑞,𝑀𝑀 , 𝑐𝑐𝑘𝑘,𝑞𝑞,𝑀𝑀 which are then
ak,q , bk,q,M , ck,q,M which are then
compared with the input error signal, where the initial input error
compared with the input error signal, where the initial input error signal is the signal
reference
is the reference
polynomial coefficients. The NI-RFmx gathers the data samples of
polynomial coefficients. The NI-RFmx gathers the data samples of both both 𝑧𝑧̂ ( n) and z(n). The
ẑ(n) and z(n). The
difference between the coefficients
difference between is thestored againisasstored
coefficients the error
again as thee(n).
signal errorThe process
signal goesprocess goes
e(n). The
on until the output
on error signal
until the reaches
output errorzero.
signalInreaches
most cases, themost
zero. In output cases, error
the signal
output reaches
error signal reaches
zero after two iterations
zero after(ntwo
= 2)iterations
[2]. Thus, (nthe
= 2)GMP DPDthe
[2]. Thus, linearizes
GMP DPD thelinearizes
PA. the PA.

Instrument Handle In Instrument Handle Out

SpecAn
Memory Polynomial Order
Reference Waveform Error Out
Error In GMP DPD

Lead Depth Lag Depth

and Order and Order

Figure 3.ofPalette
Figure 3. Palette template the GMPtemplate
DPDof the GMP
using NI’s DPD using [2].
LabVIEW NI’s LabVIEW [2].

3. Mm-Wave DPD Measurement Results on 4G LTE 64-QAM Modulated Signals

3. Mm-Wave DPD Measurement Results on 4G LTE 64-QAM Modulated Signals
We will discuss in this section the DPD measurement results on the mm-Wave two-
We will discuss
stageinCS/two-stack
this section the
GaNDPD measurement
PA [9]. results on
This is a broadband the mm-Wave
two-stage CS/2-stack two-GaN PA with
stage CS/two-stack GaN PA [9].
a small-signal This
3 dB is a broadband
bandwidth two-stage
of 19.9–35.1 GHz CS/2-stack
(55.3%), whileGaN itsPA
largewith a POUT, SAT
signal
small-signal 3dB1bandwidth
dB bandwidthof 19.9–35.1 GHz (55.3%),
may be significantly larger.while its large
The core die size of thePPA
signal OUT,is 1dB
1.9
SAT mm2 (without
bandwidth may pads),
be significantly larger.
the gain is >~20 dB,The
and core
POUT,die
MAX size
= 22ofdBm.
the PAThe is 1.9 mm2schematics,
simplified (without layout and
pads), the gain isdie 20 dB, and P
>~ micrograph ofOUT, PA=are
theMAX 22 shown
dBm. The simplified
in Figure 4. Forschematics,
the DPD work, layout and the effects
to study
on the modulated signal bandwidth, we use 1 × 20 MHz,
die micrograph of the PA are shown in Figure 4. For the DPD work, to study the effects 2 × 20 MHz, 4 × 20 MHz, and
5 × 20 MHz LTE 64-QAM signal with PAPR = 8 dB as the input signal.
on the modulated signal bandwidth, we use 1 × 20 MHz, 2 × 20 MHz, 4 × 20 MHz, and
The two-stage CS/2-stack GaN PA has a CS driver first stage with negative resistive
5 × 20 MHz LTE 64-QAM signal with PAPR = 8dB as the input signal.
feedback that drives the two-stack second stage, which is biased at a current density of
~0.2 mA/µm. For this work, the PA is operated at drain voltages (VD1 /VD2 ) of 6 V/12 V,
and the gate voltages VG1 /VG21 /VG22 are 0.2 V/0.2 V/5.8 V.
Figure 5 shows the gain and the average POUT vs. the average PIN measurement results
of the two-stage GaN PA using the 5 × 20 MHz LTE 64-QAM input signal of PAPR = 8 dB.
The PA has the highest gain improvement of 0.6 dB using the GMP DPD, with the average
PIN of around −1 dBm at 24 GHz. The gain at a higher PIN improves/expands significantly
after using the GMP DPD, as can be seen in the zoom-up inserts in Figure 5. The average
POUT improvements are highlighted and zoomed into in Figure 5, and this parameter has
the best improvement at 24 GHz. Figure 5i shows that the GMP DPD linearizes the PA
even at a lower input power (average PIN = −14 dBm) if needed. The PA has achieved a
(i) (ii) PIN of around 2 dBm/5 dBm/7 dBm at
gain improvement around 0.5 dB with an average
28 GHz/37 GHz/39 GHz using the 5 × 20 MHz LTE 64-QAM modulated input signal of
PAPR = 8 dB.
 Electronics 2023, 12, x FOR PEER REVIEW 7 of 20

Electronics 2023, 12, x FOR PEER REVIEW 7 of 20

Electronics 2023, 12, 2869 7 of 19

(iii)
Figure 4. (i) Schematic; (ii) layout; and (iii) die micrograph of the two-stage CS/2-stack broadband
GaN monolithic PA designed in our lab, which is used to perform the mm-Wave DPD measurement
results presented in this paper.

(i) The two-stage CS/2-stack GaN PA has a CS driver first (ii)stage with negative resistive
feedback that drives the two-stack second stage, which is biased at a current density of
~0.2 mA/µm. For this work, the PA is operated at drain voltages (VD1/VD2) of 6 V/12 V, and
the gate voltages VG1/VG21/VG22 are 0.2 V/0.2 V/5.8 V.
Figure 5 shows the gain and the average POUT vs. the average PIN measurement results
of the two-stage GaN PA using the 5 × 20 MHz LTE 64-QAM input signal of PAPR = 8
dB. The PA has the highest gain improvement of 0.6 dB using the GMP DPD, with the
average PIN of around −1 dBm at 24 GHz. The gain at a higher PIN improves/expands sig-
nificantly after using the GMP DPD, as can be seen in the zoom-up inserts in Figure 5. The
average POUT improvements are highlighted and zoomed into in Figure 5, and this param-
(iii)
eter has the best improvement at 24 GHz. Figure 5i shows that the GMP DPD linearizes
the PA(i)even
4. (i) at a lower and (iii) diePmicrograph
(ii)input power
and(average IN = −14 dBm)of if the
needed. The PACS/2-stack
has achieved
Figure Schematic; (ii) layout; of the two-stage CS/2-stack broadband
Figure 4. Schematic; layout; (iii) die micrograph two-stage broadband
a GaN
gain monolithic
improvement PA designed
around in our
0.5 dBlab, which
with an is used
average toPperform the mm-Wave DPD measurement
IN of around 2 dBm/5 dBm/7 dBm at
GaN monolithic
results presentedPA designed in our lab, which is used to perform the mm-Wave DPD measurement
in this paper.
28 GHz/37 GHz/39 GHz using the 5 × 20 MHz LTE 64-QAM modulated input signal of
results presented in this paper.
PAPR = 8 dB.
The two-stage CS/2-stack GaN PA has a CS driver first stage with negative resistive
feedback that drives the two-stack second stage, which is biased at a current density of
~0.2 mA/µm. For this work, the PA is operated at drain voltages (VD1/VD2) of 6 V/12 V, and
the gate voltages VG1/VG21/VG22 are 0.2 V/0.2 V/5.8 V.
Figure 5 shows the gain and 21 the average POUT vs. the average 21PIN measurement results
21 of the two-stage GaN 21 PA using the 5 × 20 MHz LTE 64-QAM input signal of PAPR = 8
20
20 dB. The PA has the highest gain improvement of 0.6 dB using17the GMP DPD, with the
17
average PIN of around −1 dBm at1924 GHz. The gain at a higher PIN improves/expands sig- POUT (dBm)
Gain (dB)

19 13
POUT (dBm)
Gain (dB)

nificantly after using the 13 GMP DPD, 18

as can be seen in the zoom-up inserts in Figure 5. The
18 average POUT improvements are highlighted and zoomed into in9Figure 5, and this param-
eter has the best improvement 9 17
at 24 GHz. Figure 5i shows that the GMP DPD linearizes
17
the PA even at a lower5input power 16 (average PIN = −14 dBm) if needed. 5 The PA has achieved
16
a gain improvement around 0.5 dB with an average PIN of around 2 dBm/5 dBm/7 dBm at
15 28 GHz/37 GHz/39 GHz 1 using 15 the 5 × 20 MHz LTE 64-QAM modulated 1 input signal of
-20 -17 PAPR
-14 -11 =-88 dB.
-5 -2 1 4 -20 -17 -14 -11 -8 -5 -2 1 4
PIN (dBm) PIN (dBm)
Electronics 2023, 12, x FOR PEERGain
REVIEW
w/oDPD Gain wDPD Gain w/oDPD Gain wDPD 8 of 20
Pout w/oDPD Pout wDPD Pout w/oDPD Pout wDPD

(i) (ii)

19 21
18 21
18 17 17
17
17 16
POUT (dBm)
Gain (dB)

13
POUT (dBm)
Gain (dB)

15 13
16
9 14
15 9
13
14 5 5
12
13 1 11 1
-20 -16 -12 -8 -4 0 4 8 -17 -13 -9 -5 -1 3 7 11
PIN (dBm) PIN (dBm)
Gain w/oDPD Gain wDPD Gain w/oDPD Gain wDPD
Pout w/oDPD Pout wDPD Pout w/oDPD Pout wDPD

(iii) (iv)
Figure 5. Measurement results of gain and POUT vs. PIN on the 2-stage CS/2-stack mm-Wave GaN PA
Figure 5. Measurement results of gain and POUT vs. PIN on the 2-stage CS/2-stack mm-Wave GaN
with/without the GMP DPD. The 64-QAM 4G LTE input signal is of 5 × 20 MHz = 100 MHz and
PA with/without
PAPR = 8 dB at the GMP
(i) 24 GHz;DPD.
(ii) 28The 64-QAM
GHz; 4G LTE
(iii) 37 GHz; GHz.signal is of 5 × 20 MHz = 100 MHz and
input
(iv) 39
PAPR = 8 dB at (i) 24 GHz; (ii) 28 GHz; (iii) 37 GHz; (iv) 39 GHz.
The power-added efficiency (PAE) of the PA has the highest improvement of only
around 2% when the PA is operated with a high average POUT of around 20 dBm at 39 GHz
using a 5 × 20 MHz LTE 64-QAM signal with PAPR = 8 dB. Figure 6 shows the compari-
son between PAEAVG and ACLR vs. the average POUT using a 5 × 20 MHz LTE 64-QAM sig-
nal with PAPR = 8 dB. The ACLR can improve very significantly using GMP DPD. At 24
Electronics 2023, 12, 2869 8 of 19

The power-added efficiency (PAE) of the PA has the highest improvement of only
around 2% when the PA is operated with a high average POUT of around 20 dBm at 39 GHz
using a 5 × 20 MHz LTE 64-QAM signal with PAPR = 8 dB. Figure 6 shows the comparison
between PAEAVG and ACLR vs. the average POUT using a 5 × 20 MHz LTE 64-QAM
signal with PAPR = 8 dB. The ACLR can improve very significantly using GMP DPD.
At 24 GHz, ACLR with DPD has a consistent improvement of around 16 dB when the
PA is operated with average POUT = 14 dBm to 16 dBm. The ACLR with GMP DPD
obtains the best improvement of around 22 dB when the PA is operated with average
POUT = 11 dBm at 39 GHz. At the common average POUT of 15 dBm/20 dBm, this PA has an
ACLR/PAE improvement of around 17 dB/0.3% using the 5 × 20 MHz LTE 64-QAM signal
at 24 GHz. At 28 GHz and 37 GHz, there is a significant ACLR/PAE improvement of around
21 dB/1.3% and 21.5 dB/1.5%, operated with a common average POUT of 13 dBm/20 dBm
for both frequencies. The PAE has the highest improvement when the PA is operated with
an average POUT of 20 dBm at 24 GHz/28 GHz/37 GHz/39 GHz.
q
EV M PA = EV Mout 2 − EV Min 2 − EV Mdriver 2 (11)

The EVM of the PA (i.e., EV M PA ) is determined from Equation (11). The intrinsic EVM
( EV Min ) is determined by having a thru connection between the two mm-Wave radio head
5582 devices as shown in Figure 2. EV Mdriver is determined by just connecting the driver
in between the mm-Wave radio head 5582 devices and measured across the similar powers
of the PA. EV Mout is the EVM from the mm-Wave DPD experiment shown in Figure 2.
Substituting each in Equation (11) gives the EV M PA .
The two-stage GaN PA after GMP DPD has an EVM that is reduced from −15.64 dBc to
−32.65 dBc (an improvement of around 17 dB), with an improved ACLR of around −29 dBc
using the 5 × 20 MHz LTE 64-QAM input of PAPR = 8 dB at 24 GHz. With an ACLR of
around—30 dBc, the PA after DPD has an EVM reduction of around 15.5 dB at 28 GHz.
As shown in Figure 7, the EVM of the PA has a major reduction when the ACLR is from
around −32 dBc to −28 dBc for both 24 GHz and 28 GHz. Figure 8 showcases the EVM
reduction in the PA across signals of different bandwidths as modulated signals. The PA is
tested using a 1 × 20 MHz, 2 × 20 MHz 4 × 20 MHz and 5 × 20 MHz LTE 64-QAM input of
PAPR = 8 dB at 28 GHz. The PA after DPD has a very large EVM reduction of around 25 dB,
with an ACLR of around −34 dBc using a 4 × 20 MHz 64-QAM 4G LTE input at 28 GHz.
This highlights that the GMP DPD works well on both narrower-band (1 × 20 MHz) and
wider-band (5 × 20 = 100 MHz) LTE 64-QAM signals. The EVM consistently has a reduction
of greater than 10 dB using the 1 × 20 MHz LTE 64-QAM input of PAPR = 8 dB at 28 GHz.
For the 2 × 20 MHz LTE 64-QAM input at 28 GHz, the EVM of the PA after DPD reduces
from −19 dBc to around −38 dBc, with an ACLR of ~−30 dBc. The rms (root-mean-square)
EVM and ACLR of the PA shown in Figure 8 are measured from a common range of average
POUT = 1 dBm to 20 dBm. With the common x-axis scale for the ACLR ranging from −48 dBc
to −20 dBc, we could see that before DPD is applied, the measured rms EVM data are
pretty much linearly proportional to the corresponding measured ACLR data, regardless of
the signal BW. However, once GMP DPD is applied, the EVM is reduced considerably at
the same level of the ACLR, and the EVM and ACLR are no longer linearly correlated for a
large range of ACLR values. For example, at the common ACLR of −34 dBc, the measured
EVM values before DPD for the 1 × 20 MHz/2 × 20 MHz/4 × 20 MHz/5 × 20 MHz
signal BW are around −23 dBc/−23 dBc/−20 dBc/−21 dBc, while after the GMP DPD
they become greatly reduced to around −37/−30/−46/−28 dBc. As bandwidth increases
from 1 × 20 MHz to 5 × 20 MHz, the rms EVM plot moves to a higher ACLR (−20 dBc)
region on the x-axes in general. The more precise BW effects on ACLR and rms EVM at
higher average output powers such as 15 dBm and 20 dBm can be seen in later figures which
are to be discussed later.
Electronics 2023, 12, x FOR PEER REVIEW 9 of 20
Electronics 2023, 12, 2869 9 of 19

15 -20 15 -20

12 -25 12 -27

ACLR (dBc)
-30

ACLR (dBc)
PAEAVG (%)
PAEAVG (%)

9 9 -34
-35
6 6 -41
-40
3 -45 3 -48

0 -50 0 -55
1 5 9 13 17 21 1 5 9 13 17 21
POUT (dBm) POUT (dBm)
PAE w/oDPD PAE wDPD PAE w/oDPD PAE wDPD
ACLR w/oDPD ACLR wDPD ACLR w/oDPD ACLR wDPD
(i) (ii)

15 -18 15 -18

-25 12 -25
12
-32 -32

ACLR (dBc)
ACLR (dBc)

PAEAVG (%)
PAEAVG (%)

9 9
-39 -39
6 6
-46 -46
3 -53 3 -53

0 -60 0 -60
1 5 9 13 17 21 1 5 9 13 17 21
POUT (dBm) POUT (dBm)
PAE w/oDPD PAE wDPD PAE w/oDPD PAE wDPD
ACLR w/oDPD ACLR wDPD ACLR w/oDPD ACLR wDPD
(iii) (iv)
Figure
Figure 6.
6. Measurement results of PAEAVG and
of PAE ACLR vs. the average POUT on the 2-stage CS/2-stack
Measurement results AVG and ACLR vs. the average POUT on the 2-stage CS/2-stack
GaN PA before/after GMP DPD using a 5 × 20 MHz 64-QAM 4G LTE input of PAPR = 8 dB at (i)
GaN PA before/after GMP DPD using a 5 × 20 MHz 64-QAM 4G LTE input of PAPR = 8 dB at
24 GHz; (ii) 28 GHz; (iii) 37 GHz; and (iv) 39 GHz.
(i) 24 GHz; (ii) 28 GHz; (iii) 37 GHz; and (iv) 39 GHz.
The two-stage GaN PA after GMP DPD has an EVM that is reduced from −15.64 dBc
to −32.65 dBc (an improvement of around 17 dB), with an improved ACLR of around −29
dBc using the 5 × 20 MHz LTE 64-QAM input of PAPR = 8 dB at 24 GHz. With an ACLR
of around—30 dBc, the PA after DPD has an EVM reduction of around 15.5 dB at 28 GHz.
As shown in Figure 7, the EVM of the PA has a major reduction when the ACLR is from
around −32 dBc to −28 dBc for both 24 GHz and 28 GHz. Figure 8 showcases the EVM
reduction in the PA across signals of different bandwidths as modulated signals. The PA
 × 20 MHz/4 × 20 MHz/5 × 20 MHz signal BW are around −23 dBc/−23 dBc/−20 dBc/−21
dBc, while after the GMP DPD they become greatly reduced to around −37/−30/−46/−28
dBc. As bandwidth increases from 1 × 20 MHz to 5 × 20 MHz, the rms EVM plot moves
to a higher ACLR (−20 dBc) region on the x-axes in general. The more precise BW effects
Electronics 2023, 12, 2869
on ACLR and rms EVM at higher average output powers such as 15 dBm and 20 dBm can
10 of 19
be seen in later figures which are to be discussed later.

-6 -4
-11 -10
rms EVM (dB)

rms EVM (dB)

-16 -16

-21 -22

-26 -28

-31 -34

-36 -40
-40 -36 -32 -28 -24 -20 -46 -42 -38 -34 -30 -26 -22
ACLR (dBc) ACLR (dBc)
EVM w/oDPD EVM wDPD EVM w/oDPD EVM wDPD

(i) (ii)
Figure 7. Measurement results of the rms EVM vs. ACLR on the 2-stage CS/2-stack GaN PA be-
Figure 7. Measurement results of the rms EVM vs. ACLR on the 2-stage CS/2-stack GaN PA
fore/after GMP DPD using 5 × 20 MHz 64-QAM 4G LTE input of PAPR = 8 dB at: (i) 24 GHz; (ii)
before/after GMP DPD using 5 × 20 MHz 64-QAM 4G LTE input of PAPR = 8 dB at: (i) 24 GHz;
28 GHz.
(ii) 28 GHz.

The measured constellation diagram and AM/gain distortion of the PA are shown
in Figure 9. The PA is tested using the 5 × 20 MHz LTE 64-QAM modulated signal
input of PAPR = 8 dB at 24 GHz/28 GHz/37 GHz/39 GHz with/without the GMP DPD.
Figure 9(I) presents the measured rms EVM, ACLR and PAEAVG vs. the PA’s average
POUT at relatively higher output power levels (i.e., average POUT = 6 dBm to 20 dBm) at
a given mm-Wave frequency (four plots, (i)–(iv), at 24 GHz/28 GHz/37 GHz/39 GHz,
respectively). Figure 9(i–iv)(a,b) shows the PA transmit output spectrum, constellation
diagram and AM/gain distortion curve when the PA is operated at average POUT ~12 dBm
at a given frequency ((i)–(iv), 24–39 GHz, using the 5 × 20 MHz LTE 64-QAM signal input
of PAPR = 8 dB). At 24 GHz, with an average POUT ~12 dBm and after applying the GMP
DPD, the GaN PA has a rms EVM reduction from −20.5 dBc to −27.5 dBc, and an ACLR
improvement of around 15 dB. The non-linear AM/fain distortion curves are plotted with
an input power of −59 dBm to −38 dBm on the horizontal axis, which corresponds to an
absolute PIN of −18 dBm to 3 dBm due to the 41 dB attenuation used in the setup for 24 GHz.
The PA output spectrum at 24 GHz shows that its ACLR +/− is −33.6 dBc/−33.4 dBc
before the GMP DPD, and after the GMP DPD a significant reduction in the ACLR +/− of
−48.4 dBc/−45.6 dBc is observed. We can also see from the constellation diagram before
the DPD, that the symbols are more scattered and rotated (in red), but after the GMP DPD,
these symbols are slightly better aligned and centered at POUT ~12 dBm (in blue). The
relationship of ACLR, rms EVM and PAEAVG is also plotted against the average POUT at
each frequency in Figure 9, which shows that at a higher average POUT , the PA’s linearity
will inevitably degrade with a higher ACLR and rms EVM, but the PA will also have higher
efficiency as it is more saturated (e.g., PAEAVG ~11% to 14%) with a larger average POUT (e.g.,
18 dBm to 20 dBm). At these higher average POUT levels of 18–20 dBm, the corresponding
peak POUT is close to the Watt level (i.e., 26–28 dBm), and the rms EVM values are around
−18 dB to −12 dB and ACLR~−40 dBc to −32 dBc without DPD. After the GMP DPD is
applied, the rms EVM and ACLR can be improved by ~10 dB and 3–4 dB, respectively,
at these high POUT levels where the PA is a bit more nonlinear. From the 4–6 dB power
back-off levels at 24 GHz where the average POUT is 14 dBm to 16 dBm, the PA can have an
Electronics 2023, 12, 2869 11 of 19

excellent ACLR of ~−46 dBc after the GMP DPD, where the rms EVM and ACLR can both
be improved by ~15 dB or more, with a PAEAVG of 4% to 6%. Similarly, Figure 9 shows
that at 28 GHz/37 GHz/39 GHz, when the average POUT is from 13 dBm to 15 dBm, 11
Electronics 2023, 12, x FOR PEER REVIEW theofPA
20
has an outstanding ACLR from −50 dBc to −46 dBc, −52 dBc to −46 dBc, and −52 dBc to
−46 dBc, respectively.

-18 -4

-24 -10
rms EVM (dB)

rms EVM (dB)

-16
-30
-22
-36
-28
-42 -34

-48 -40
-48 -41 -34 -27 -20 -48 -41 -34 -27 -20
ACLR (dBc) ACLR (dBc)

EVM w/oDPD EVM wDPD EVM w/oDPD EVM wDPD

(i) (ii)

-10 -4
-16 -10
rms EVM (dB)

rms EVM (dB)

-22 -16

-28 -22

-34 -28

-40 -34

-46 -40
-48 -41 -34 -27 -20 -48 -41 -34 -27 -20
ACLR (dBc) ACLR (dBc)

EVM w/oDPD EVM wDPD EVM w/oDPD EVM wDPD

(iii) (iv)
Figure 8. Measurement results of the rms EVM vs. ACLR on 2-stage CS/2-Stack GaN PA before/after
Figure 8. Measurement results of the rms EVM vs. ACLR on 2-stage CS/2-Stack GaN PA before/after
the GMP DPD using (i) 1 × 20 MHz; (ii) 2 × 20 MHz; (iii) 4 × 20 MHz; (iv) 5 × 20 MHz 64-QAM
the GMP DPD using (i) 1 × 20 MHz; (ii) 2 × 20 MHz; (iii) 4 × 20 MHz; (iv) 5 × 20 MHz 64-QAM 4G
4G LTE input signals of PAPR = 8 dB at 28 GHz.
LTE input signals of PAPR = 8 dB at 28 GHz.
The measured constellation diagram and AM/gain distortion of the PA are shown in
Figure 9. The PA is tested using the 5 × 20 MHz LTE 64-QAM modulated signal input of
PAPR = 8 dB at 24 GHz/28 GHz/37 GHz/39 GHz with/without the GMP DPD. Figure 9(I)
presents the measured rms EVM, ACLR and PAEAVG vs. the PA’s average POUT at relatively
higher output power levels (i.e., average POUT = 6 dBm to 20 dBm) at a given mm-Wave
 ACLR can be improved by ~10 dB and 3–4 dB, respectively, at these high POUT levels where
the PA is a bit more nonlinear. From the 4–6 dB power back-off levels at 24 GHz where
the average POUT is 14 dBm to 16 dBm, the PA can have an excellent ACLR of ~−46 dBc after
the GMP DPD, where the rms EVM and ACLR can both be improved by ~15 dB or more,
with a PAEAVG of 4% to 6%. Similarly, Figure 9 shows that at 28 GHz/37 GHz/39 GHz, when
Electronics 2023, 12, 2869 the average POUT is from 13 dBm to 15 dBm, the PA has an outstanding ACLR from −5012dBc
of 19
to −46 dBc, −52 dBc to −46 dBc, and −52 dBc to −46 dBc, respectively.

(a) (b) (c)

-4 15
(I)

ACLR (dBc), rms EVM (dB)

12
-18

PAEAVG (%)
9
-32
(II)
6

-46
3
No DPD: with DPD

Pout ~ 12.1dBm : 12.2dBm

-60 0
rms EVM ~ -20.5dB : -27.5dB 5 9 13 17 21
PAEavg ~ 2.6% : 2.6% POUT (dBm)
ACLR ~ -33.6/-33.4dBc : -48.4/-45.6dBc
ACLR w/oDPD ACLR wDPD EVM w/oDPD
Constellation Plots (I) Adjacent Channel Leakage Power Ratio (ACLR)
Electronics 2023, 12, x FOR PEER REVIEW (II) AM/Gain Distortion EVM wDPD PAE w/oDPD 13 of 20
PAE wDPD

(i)

(a) (b) (c)

-4 15
(I)
ACLR (dBc), rms EVM (dB)

12
-18

PAEAVG (%)
9
-32
(II) 6

-46
3
No DPD: with DPD

-60 0
Pout ~ 12.6dBm : 12.6dBm
rms EVM ~ -20.5dB : -28.3dB
5 9 13 17 21
PAEavg ~ 2.7% : 2.8% POUT (dBm)
ACLR ~ -33.4/-33.5dBc : -53.9/-54.1dBc
Constellation Plots (I) Adjacent Channel Leakage Power Ratio (ACLR) ACLR w/oDPD ACLR wDPD EVM w/oDPD
(II) AM/Gain Distortion
EVM wDPD PAE w/oDPD PAE wDPD

(ii)
(a) (b) (c)
Figure 9. Cont.
-4 15
(I)
ACLR (dBc), rms EVM (dB)

12
-18
PAEAVG (%)

9
-32
(II)
6

-46
3
No DPD: with DPD

Pout ~ 12.6dBm : 12.7dBm

-60 0
rms EVM ~ -18.4dB : -38.1dB 5 9 13 17 21
 No DPD: with DPD

-60 0
Pout ~ 12.6dBm : 12.6dBm
rms EVM ~ -20.5dB : -28.3dB
5 9 13 17 21
PAEavg ~ 2.7% : 2.8% POUT (dBm)
ACLR ~ -33.4/-33.5dBc : -53.9/-54.1dBc
Constellation
Electronics 2023, 12, 2869 Plots (I) Adjacent Channel Leakage Power Ratio (ACLR) ACLR w/oDPD ACLR wDPD EVM w/oDPD
13 of 19
(II) AM/Gain Distortion
EVM wDPD PAE w/oDPD PAE wDPD

(ii)
(a) (b) (c)
-4 15
(I)

ACLR (dBc), rms EVM (dB)

12
-18

PAEAVG (%)
9
-32
(II)
6

-46
3
No DPD: with DPD

Pout ~ 12.6dBm : 12.7dBm

-60 0
rms EVM ~ -18.4dB : -38.1dB 5 9 13 17 21
PAEavg ~ 2.7% : 2.8% POUT (dBm)
ACLR+/- ~ -31.6/-31.4dBc : -49.5/-52.9dBc
ACLR w/oDPD ACLR wDPD EVM w/oDPD
Electronics 2023,Constellation (I) Adjacent Channel Leakage Power Ratio (ACLR)
Plots REVIEW
12, x FOR PEER 14 of 20
(II) AM/Gain Distortion EVM wDPD PAE w/oDPD PAE wDPD

(iii)
(a) (b) (c)
-4 15
(I)
ACLR (dBc), rms EVM (dB)

12
-18

PAEAVG (%)
9
-32
(II) 6

-46
3
No DPD: with DPD

-60 0
Pout ~ 12.4dBm : 12.5dBm
rms EVM ~ -20.1dB : -30.0dB 5 9 13 17 21
PAEavg ~ 2.5% : 2.5% POUT (dBm)
ACLR ~ -33.2/-32.9dBc : -55.4/-54.5dBc
(I) Adjacent Channel Leakage Power Ratio (ACLR) ACLR w/oDPD ACLR wDPD EVM w/oDPD
Constellation Plots
(II) AM/Gain Distortion EVM wDPD PAE w/oDPD PAE wDPD

(iv)
Figure 9. Measurement results of the 2-stage CS/2-stack GaN PA (a) constellation diagrams; (b) PA
Figure 9. Measurement results of the 2-stage CS/2-stack GaN PA (a) constellation diagrams; (b) PA
output spectra, gain vs. PIN; and (c) the ACLR, rms EVM and PAEAVG that are also plotted against the
output spectra, gain vs. PIN ; and (c) the ACLR, rms EVM and PAEAVG that are also plotted against
average the POUT of the PA before/after GMP DPD using a 5 × 20 MHz 64-QAM 4G LTE input signal
the average the POUT of the PA before/after GMP DPD using a 5 × 20 MHz 64-QAM 4G LTE input
of PAPR = 8 dB at (i) 24 GHz; (ii) 28 GHz; (iii) 37 GHz; (iv) 39 GHz.
signal of PAPR = 8 dB at (i) 24 GHz; (ii) 28 GHz; (iii) 37 GHz; (iv) 39 GHz.
As also seen from Figure 9ii(b)–iv(b), the PA transmit output spectra before/after the
GMP DPD show that the PA has a rms EVM reduction of from −20.5 dB to −28.3 dB/−18.4
dB to −38.1 dB/−20.1 dB to −30.0 dB, an ACLR improvement of around 21 dB/21 dB/22 dB,
and a PAE improvement of around 0.1%/0.1%/0.0% with GMP DPD at 28 GHz/37 GHz/39
GHz, respectively. We also see that the PA transmit output spectrum after GMP DPD has
an ACLR+/− of −53.9 dBc/−54.1 dBc and −55.4 dBc/−54.5 dBc at 28 GHz and 39 GHz, re-
spectively. These good linearity data suggest that the PA memory effects are considerably
reduced after using GMP DPD. The AM/fain non-linear distortion curve highlighted in
red is shown in Figure 9 as well. For 28 GHz/37 GHz/39 GHz, the PAs are linear as shown
in blue highlights scaling from an input power of −40 dBm to −32 dBm/−32 dBm to −24
dBm/−32 dBm to −26 dBm at their respective frequencies. In Figure 9i(b)–iv(b), the non-
linearities of the PA transmit output spectra across the adjacent channels (in red) at a given
Electronics 2023, 12, 2869 14 of 19

As also seen from Figure 9ii(b)–iv(b), the PA transmit output spectra before/after the
GMP DPD show that the PA has a rms EVM reduction of from −20.5 dB to −28.3 dB/−18.4 dB
to −38.1 dB/−20.1 dB to −30.0 dB, an ACLR improvement of around 21 dB/21 dB/22 dB, and
a PAE improvement of around 0.1%/0.1%/0.0% with GMP DPD at 28 GHz/37 GHz/39 GHz,
respectively. We also see that the PA transmit output spectrum after GMP DPD has an
ACLR+/− of −53.9 dBc/−54.1 dBc and −55.4 dBc/−54.5 dBc at 28 GHz and 39 GHz,
respectively. These good linearity data suggest that the PA memory effects are considerably
reduced after using GMP DPD. The AM/fain non-linear distortion curve highlighted in
red is shown in Figure 9 as well. For 28 GHz/37 GHz/39 GHz, the PAs are linear as shown
in blue highlights scaling from an input power of −40 dBm to −32 dBm/−32 dBm to
−24 dBm/−32 dBm to −26 dBm at their respective frequencies. In Figure 9i(b)–iv(b), the
non-linearities of the PA transmit output spectra across the adjacent channels (in red) at a
given center frequency (i.e., 24 GHz/28 GHz/37 GHz/39 GHz) are further worsened due
to CA of the input modulated signal with 5 × 20 MHz, as similarly reported in Refs. [21,22].
This ACLR non-linearity increases as the operating frequency rises and is considerably
reduced using GMP DPD (in blue). For example, looking at the adjacent channels of
Figure 9 at 24 GHz (from 24.05 GHz to 24.15 GHz) with the BW of 5 × 20 MHz, we see that
the ACLR jumps (in red) become a bit further reduced after applying GMP DPD (in blue),
and they increase as the frequency increases to 28 GHz/37 GHz/39 GHz.
Figure 10 demonstrates the ACLR improvement across various frequencies and average
POUT values using a 5 × 20 MHz LTE 64-QAM input signal of PAPR = 8 dB. The PA after
GMP DPD has an excellent ACLR improvement for the lower average POUT levels (say,
an average POUT of ≤ 13 dBm) at all frequencies, while the best ACLR happens at the
highest frequency of 39 GHz and the ACLR worsens as frequencies become lower, reaching
the worst ACLR at 24 GHz. The measured ACLR reaches around −59 dBc at 39 GHz at
the average POUT of 9 dBm, which is the best ACLR level for all cases presented in this
work. However, as the PA is operated at a higher average POUT from 15 dBm to 21 dBm,
the lowest ACLR now happens at the lowest frequency of 24 GHz, and worsens as the
frequencies become higher, reaching the worst ACLR at the highest testing frequency of
39 GHz. Therefore, this trend of reaching the best ACLR over the frequencies depends
on the average POUT level, and interestingly it changes around the average POUT level of
around 13 dBm, suggesting that it would be very challenging to use the same GMP DPD
algorithm to produce a broadband PA [23,24] linearity improvement consistently with
a large average POUT level and frequency range. The linearity improvement is mostly
independent of the operation frequency of the broadband mm-Wave PAs tested in this
work, which is encouraging as this demonstrates the effectiveness of the GMP DPD model
used. The bandwidth effects of the LTE 64-QAM input signal on the PA’s ACLR are shown
in Figure 11 at the single frequency of 28 GHz. We can see that the PA’s ACLR after GMP
DPD using a 1 × 20 MHz 64-QAM 4G LTE input of PAPR = 8 dB is better than those from
the 2 × 20 MHz/4 × 20 MHz/5 × 20 MHz LTE 64-QAM input. It appears that the wider the
modulated signal BW, the worse the ACLR becomes. However, evidently when the PA is
operated with an average POUT of 13 dBm, the ACLR after GMP DPD the using 1 × 20 MHz
and 5 × 20 MHz LTE 64-QAM input of PAPR = 8 dB modulated signals at 28 GHz are
equally around -54 dBc. As discussed earlier, we can see that the ACLR decreases from
−30 dBc to −43 dBc as the signal bandwidth decreases from 5 × 20 MHz to 1 × 20 MHz
at an average POUT of 20 dBm, which describes the BW effect on the ACLR. This BW effect
similarly occurs with the lower input power as well. The trend changes for the POUT from
9 dBm to 15 dBm as the ACLR of 5 × 20 MHz is greater than the ACLR of 2 × 20 MHz and
4 × 20 MHz at 28 GHz are due to the GMP DPD.
 dBc. As discussed earlier, we can see that the ACLR decreases from −30 dBc to −43 dBc as
the signal bandwidth decreases from 5 × 20 MHz to 1 × 20 MHz at an average POUT of 20
dBm, which describes the BW effect on the ACLR. This BW effect similarly occurs with
the lower input power as well. The trend changes for the POUT from 9 dBm to 15 dBm as
Electronics 2023, 12, 2869 the ACLR of 5 × 20 MHz is greater than the ACLR of 2 × 20 MHz and 4 × 20 MHz 15 ofat
19
28 GHz are due to the GMP DPD.

-25

-32

ACLR (dBc)
-39

-46

-53

-60
1 5 9 13 17 21
POUT (dBm)
ACLR wDPD @24GHz ACLR wDPD @28GHz
ACLR wDPD @37GHz ACLR wDPD @39GHz

Figure 10. Measurement results of ACLR vs. the average POUT on the 2-stage CS/2-stack GaN PA after
Figure 10. Measurement results of ACLR vs. the average POUT on the 2-stage CS/2-stack GaN PA
Electronics 2023, 12, x FOR PEER REVIEW
DPD using a 5 × 20 MHz 64-QAM 4G LTE input signal of PAPR = 8 dB at (i) 24 GHz; (ii) 2816GHz;
of 20
after DPD using a 5 × 20 MHz 64-QAM 4G LTE input signal of PAPR = 8 dB at (i) 24 GHz; (ii) 28 GHz;
(iii) 37 GHz; (iv) 39 GHz.
(iii) 37 GHz; (iv) 39 GHz.

-25
1 × 20 MHz 2 × 20 MHz
4 × 20 MHz 5 × 20 MHz
-32
ACLR (dBc)

-39

-46

-53

-60
1 5 9 13 17 21
POUT (dBm)
Figure 11.
Figure 11. Measurement
Measurement results
results of
ofACLR
ACLRvs. averagePPOUT
vs.average on the
OUT on the 2-stage
2-stage CS/2-stack
CS/2-stack GaN PA after
GaN PA after
DPD using (i) 1 × 20 MHz; (ii) 2 × 20 MHz; (iii) 4 × 20 MHz; (iv) 5 × 20 MHz 64-QAM
DPD using (i) 1 × 20 MHz; (ii) 2 × 20 MHz; (iii) 4 × 20 MHz; (iv) 5 × 20 MHz 64-QAM 4G LTE 4Ginput
LTE
input signal of PAPR = 8 dB at
signal of PAPR = 8 dB at 28 GHz. 28 GHz.

As the
As the PA
PA isisoperated
operatedat atan averagePPOUT
anaverage OUT greater
greater than
than 17 dBm, the
17 dBm, the ACLR
ACLR after
after GMP
GMP
DPD becomes
DPD becomesworse
worsewith
witha alarger
larger signal
signal bandwidth.
bandwidth. TheThe ACLR
ACLR values
values after
after GMPGMPDPD DPD
are
are around
around −42 dBc/−36
−42 dBc/ −36 dBc/dBc/−32
−32 dBc/ dBc/−30
−30 dBc dBc
forfor
thethe
1 × 120×MHz/2
20 MHz/2 × 20 MHz/4
× 20 MHz/4 × 20
× 20 MHz/
5MHz/5 × 20LTE
× 20 MHz MHz LTE 64-QAM
64-QAM modulated modulated input
input signal signal operated
operated with an
with an average average
POUT POUT
of 20 of
dBm.
20 dBm.
Figure 12 shows the modulated signal bandwidth effects on the rms EVM and ACLR
acrossFigure
1 × 2012MHz/2
shows the modulated
× 20 MHz/4 × signal bandwidth
20 MHz/5 × 20 effects
MHz at onathe rms average
higher EVM and ACLR
POUT (of
≥across 1 × 20 MHz/2 × 20 MHz/4 ×OUT,
15 dBm and 20 dBm). At an average P 20 MHz/5
MAX of × 20 MHz at a higher average POUT (of ≥
15 dBm, the rms EVM/ACLR after GMP
15 dBm
DPD and 20from
improves dBm). At an
−24.3 dB/ average
−34.9 P
dBc
OUT,to of 15dB/-51.7
−44.5
MAX dBm, the rms
dBc; EVM/ACLR
from −22.6 dB/after
−32.6GMP
dBc
DPD improves from −24.3 dB/−34.9 dBc to −44.5 dB/-51.7 dBc; from −22.6 dB/−32.6 dBc to
−30.2 dB/−48.6 dBc; from −15.7 dB/−29.7 dBc to −22.2 dB/−41.9 dBc; and from −13.9 dB/−26.7
dBc to −20.0 dB/−39.1 dBc using the corresponding 1 × 20 MHz, 2 × 20 MHz, 4 × 20
MHz, and 5 × 20 MHz LTE 64-QAM input signals. When the average POUT, MAX = 20 dBm,
the rms EVM/ACLR after GMP DPD improves from −19.2 dB/−29.9 dBc to −31.2 dB/−41.5
Electronics 2023, 12, 2869 16 of 19

to −30.2 dB/−48.6 dBc; from −15.7 dB/−29.7 dBc to −22.2 dB/−41.9 dBc; and from
−13.9 dB/−26.7 dBc to −20.0 dB/−39.1 dBc using the corresponding 1 × 20 MHz,
2 × 20 MHz, 4 × 20 MHz, and 5 × 20 MHz LTE 64-QAM input signals. When the average
POUT, MAX = 20 dBm, the rms EVM/ACLR after GMP DPD improves from −19.2 dB/−29.9 dBc
to −31.2 dB/−41.5 dBc; −15.6 dB/−27.4 dBc to −22.7 dB/−37.4 dBc; −11.8 dB/−24.7 dBc to
−14.8 dB/−31.9 dBc, and −9.0 dB/−22.5 dBc to −11.8 dB/−30.2 dBc using the 1 × 20 MHz,
2 × 20 MHz, 4 × 20 MHz and 5 × 20 MHz LTE 64-QAM input signals, respectively. The
difference in the ACLR after GMP DPD decreases as the bandwidth is increased for both
cases of average POUT of 15 dBm and 20 dBm. This trend is also similar to the ACLR for
the rms EVM to reduce after DPD, and it also worsens as the signal bandwidth increases.
Wideband-modulated 5G signals are achieved using CA. The distortion worsens as the
signal BW increases, with a higher AM/AM and with the rms phase error of AM/PM
rising as the signal BW increases, as well as worsened ACLR and EVM values. However,
Electronics 2023, 12, x FOR PEER REVIEW 17 of 20
considerable linearity improvements are seen after GMP DPD, which are especially relevant
and critical to extending mm-Wave GaN PAs for mm-Wave applications that require larger
POUT, Linear values such as 5G small cells.

-12 -8
ACLR (dBc), rms EVM (dB)

ACLR (dBc), rms EVM (dB)

-20 -15

-28 -22

-36 -29

-44 -36

-52 -43
20 40 60 80 100 20 40 60 80 100
BW (MHz) BW (MHz)
ACLR w/oDPD ACLR wDPD ACLR w/oDPD ACLR wDPD
EVM w/oDPD EVM wDPD EVM w/oDPD EVM wDPD
(i) (ii)
Figure 12. Measurement results of the rms EVM and ACLR vs. BW on the 2-stage CS/2-stack GaN
Figure 12. Measurement results of the rms EVM and ACLR vs. BW on the 2-stage CS/2-stack
PA before/after DPD using the 64-QAM 4G LTE input signal of PAPR = 8 dB at 28 GHz. (i) POUT, MAX
GaN PA before/after
= 15 dBm; (ii) POUT, MAX DPD using the 64-QAM 4G LTE input signal of PAPR = 8 dB at 28 GHz.
= 20 dBm.
(i) POUT, MAX = 15 dBm; (ii) POUT, MAX = 20 dBm.
4. State-of
4. State-of Art Literature Comparison
Art Literature Comparison
This section
This section compares
comparesour ourwork
worktoto that onon
that thethe
various state-of-the-art
various mm-Wave
state-of-the-art mm-Wave PAs
using the 4G LTE modulated signals reported in the literature as shown
PAs using the 4G LTE modulated signals reported in the literature as shown in Table 1. in Table 1. Our
two-stage
Our CS/two-stack
two-stage CS/two-stack GaNGaNPA PAhashas
thethebest EVM
best EVMimprovement
improvementininthe the literature,
literature, ofof
around 18 dB after GMP DPD, using the 1 × 20 MHz LTE 64-QAM input with PAPR = 8 dB8
around 18 dB after GMP DPD, using the 1 × 20 MHz LTE 64-QAM input with PAPR =
dB and
and withwith an average
an average POUTPOUT
of of
14 14
dBm dBm
[8].[8]. Additionally,
Additionally, ourour
GaNGaN PAPA operated
operated at an
at an aver-
average
age P OUT of 15 dBm has a large ACLR improvement from −28.3 dBc to −45.6 dBc using the
POUT of 15 dBm has a large ACLR improvement from −28.3 dBc to −45.6 dBc using
5 ×520
the × MHz
20 MHz LTELTE64-QAM
64-QAM input signal
input withwith
signal PAPR = 8 dB.
PAPR = 8Among the 2 the
dB. Among × 202 MHz mod-
× 20 MHz
ulated signal literature survey, [7] has an ACLR improvement of around
modulated signal literature survey, [7] has an ACLR improvement of around 11.5 dB 11.5 dB operating
at a POUT =at
operating 14adBm,
POUT whereas
= 14 dBm, our GaN PA
whereas ourhas an improvement
GaN of around of
PA has an improvement 15.3 dB using
around 15.3 the
dB
2 × 20
using MHz
the 2 × LTE 64-QAM
20 MHz modulated
LTE 64-QAM signal with
modulated PAPR
signal with= 8PAPR
dB operating POUTat=
also at a also
= 8 dB operating
a14POUT
dBm.= In
14 [7],
dBm.theIntwo-stack/CS CMOS PA
[7], the two-stack/CS uses PA
CMOS memory polynomial
uses memory DPD with
polynomial DPD 1 ×with20
MHz/4 × 20 MHz LTE 64-QAM modulated input signals. At 28 GHz [6], its ACLR im-
provement of around 17 dB occurs for the 4 × 20 MHz modulated signals operating at an
average POUT of 11 dBm, while our GaN PA has an improvement of around 15 dB along
with an EVM reduction of around 19 dB operating at an average POUT of 11 dBm. The PA
has an ACLR improvement/EVM reduction of around 16.9 dB/17 dB, 17.3 dB/12.6 dB, 16.5
Electronics 2023, 12, 2869 17 of 19

1 × 20 MHz/4 × 20 MHz LTE 64-QAM modulated input signals. At 28 GHz [6], its ACLR
improvement of around 17 dB occurs for the 4 × 20 MHz modulated signals operating at
an average POUT of 11 dBm, while our GaN PA has an improvement of around 15 dB along
with an EVM reduction of around 19 dB operating at an average POUT of 11 dBm. The PA
has an ACLR improvement/EVM reduction of around 16.9 dB/17 dB, 17.3 dB/12.6 dB,
16.5 dB/8.3 dB, 15.1 dB/7 dB using the 5 × 20 MHz LTE 64-QAM modulated signal with
PAPR =8 dB at 24 GHz, 28 GHz, 37 GHz, and 39 GHz at an average POUT of 15 dBm,
respectively. At the large average POUT of 20 dBm, an ACLR improvement/EVM reduction
of around 8.9 dB/3.4 dB, 9.5 dB/2.9 dB, 7.7 dB/2.3 dB, 11.2 dB/2.9 dB using the 5 × 20 MHz
LTE 64-QAM modulated signal with PAPR =8 dB is seen at 24 GHz, 28 GHz, 37 GHz, and
39 GHz, respectively.

Table 1. Tabulated narrowband/broadband DPD performance comparison of state-of-the-art mm-

Wave PAs using 4G LTE modulated signals in the literature [6–8].

ACLR ACLR EVM EVM

DPD Topol- Modula- Freq. Pout, avg BW PAPR
Ref. Tech. w/oDPD wDPD w/oDPD wDPD
Work ogy tion (GHz) (dBm) (MHz) (dB)
(dBc) (dBc) (dB) (dB)
−20 −33 14 20 −28.1 - 7.5
2-Stack −20 −38 7 80 −27.3 - 10.8
MP 64-QAM
[6] CMOS 28
DPD 4G LTE −20 −33 11 20 −25.7 - 7.5
CS
−18 −35 11 80 −27.5 - 7.2
GMP - - 64-QAM −35.8 −47.3 14 20 −17.4 −34.7 7.2
[7] 27
DPD 4G LTE −33 −45 14 40 −20.3 −32.3 7.2
High
I/Q Im- GaAs
[8] - 4G LTE 28 −33 −45 - 100 - - 10.8
balance pHEMT
DPD
−29.1 −46.0 15 100 −15.6 −32.6
24
−22.8 −31.7 20 100 −11.2 −14.6
−37.8 −54.3 14 20 −27.5 −45.5
−34.4 −49.7 14 40 −23 −31
2-Stage 28 −35.7 −50.4 11 80 −21.8 −40.1
This GMP GaN 64-QAM −28.3 −45.6 15 100 −15.4 −28.0
CS/2- 8
Work DPD On SiC 4G LTE −22.0 −31.5 20 100 −9.5 −12.4
Stack
−27.7 −44.2 15 100 −13.9 −22.2
37
−22.5 −30.2 20 100 −9.0 −11.3
−27.5 −42.6 15 100 −14.6 −21.6
39
−22.9 −34.1 20 100 −9.5 −12.4

5. Conclusions
From the mm-Wave GaN PA measurement results in Section 3, we can see a significant
improvement in gain compression and performance using GMP DPD, even at higher
input power and frequencies. After implementing GMP DPD, the ACLR of the PA shows
excellent linearization at 37 GHz using 5 × 20 MHz 64-QAM 4G LTE signals. We can
conclude that the GMP DPD is effective at linearizing our two-stage CS/two-stack GaN
PA. This work shows promising results even when the two-stage GaN PA is operated at
a higher average POUT of 15 dBm and 20 dBm, as the GMP DPD significantly improves
the ACLR and EVM linearity of the two-stage CS/two-stack GaN PA using 1 × 20 MHz,
2 × 20 MHz, 4 × 20 MHz, and 5 × 20 MHz 64-QAM LTE input signals. The measured
linearity improvement and performance of our two-stage CS/two-stack GaN PA using
GMP DPD are very impressive when compared against those of all other state-of-the-
art papers in the literature as shown in Table 1, especially over such a large operational
frequency range of 24–39 GHz [6–8]. To the authors’ best knowledge, our work presented
here is the first reported paper to have successfully linearized a monolithic broadband mm-
Wave GaN PA using the GMP DPD algorithm across the key 5G FR2 band, as demonstrated
with test data using aggressive modulated waveforms (i.e., LTE 64-QAM) of various signal
bandwidths [10–13]. We would like to mention that applying GMP DPD on the mm-Wave
GaN PAs, even though it is shown in this work to provide significant tradeoffs between
clear improvements in linearity and PAE design, would also come with a price as the DPD
Electronics 2023, 12, 2869 18 of 19

conducted at the baseband also adds to the power consumption and complicates baseband
signal processing. So far, we have only used the NI testbench to characterize the PAs
with DPD algorithms as a feasibility study, but we have neither worked on evaluating the
extra power practically needed by adding DPD using a FPGA or into an IC to implement
the algorithms, nor have we investigated analog predistortion options. We would like to
explore these other linearization techniques further in the future when funding allows.

Author Contributions: Conceptualization, G.S. and D.Y.C.L.; methodology, G.S. and D.Y.C.L.; vali-
dation, G.S., C.S. and D.Y.C.L.; formal analysis, G.S.; investigation, G.S.; resources, J.C.M.; D.Y.C.L.
and J.L.; data curation, G.S. and C.S.; writing—original draft preparation, G.S.; writing—review and
editing, G.S. and D.Y.C.L.; visualization, G.S.; supervision, D.Y.C.L.; funding acquisition, D.Y.C.L. All
authors have read and agreed to the published version of the manuscript.
Funding: This material is based partly on research sponsored by the Air Force Research Laboratory
(AFRL) and the Defense Advanced Research Projects Agency (DARPA) under grant number FA8650-
19-1-7902. The US Government is authorized to reproduce and distribute reprints for governmental
purposes notwithstanding any copyright notation thereon. The authors would also to sincerely thank
the donors for the TTU Keh-Shew Lu Regents Chair Endowment fund.
Data Availability Statement: Data contained within this article.
Acknowledgments: We would like to thank Ouyang Liang-Wei for their support in this work. We
also sincerely thank the GlobalFoundries (GF) University Program, HRL, for the tape out support
and training.
Conflicts of Interest: The authors declare no conflict of interest. The funders had no role in the design
of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or
in the decision to publish the results.

References
1. Available online: www.3gpp.org/ (accessed on 27 March 2023).
2. National Instruments. Fundamentals of Power Amplifier Testing. Available online: https://www.ni.com/en-us/support/
documentation/supplemental/16/rfic-white-paper-series{-}{-}fundamentals-of-power-amplifier-testing.html (accessed on 27
March 2023).
3. Lie, D.Y.; Mayeda, J.C.; Li, Y.; Lopez, J. A review of 5G power amplifier design at cm-wave and mm-wave frequencies. Wirel.
Commun. Mob. Comput. 2018, 2018, 6793814. [CrossRef]
4. Lie, D.Y.C.; Mayeda, J.C.; Lopez, J. Highly efficient 5G linear power amplifiers (PA) design challenges. In Proceedings of the
2017 International Symposium on VLSI Design, Automation and Test (VLSI-DAT), Hsinchu, Taiwan, 24–27 April 2017; IEEE:
Piscataway, NJ, USA, 2017; pp. 1–3.
5. Somasundaram, G.; Mayeda, J.C.; Lie, D.Y.; Lopez, J. Effective Linearization of Millimeter-Wave Power Amplifiers Using
LUT-Based Digital Predistortion (DPD). In Proceedings of the 2022 IEEE Texas Symposium on Wireless and Microwave Circuits
and Systems (WMCS), Waco, TX, USA, 19–20 April 2022; IEEE: Piscataway, NJ, USA, 2022; pp. 1–4.
6. Park, B.; Jin, S.; Jeong, D.; Kim, J.; Cho, Y.; Moon, K.; Kim, B. Highly linear mm-wave CMOS power amplifier. IEEE Trans. Microw.
Theory Tech. 2016, 64, 4535–4544. [CrossRef]
7. Yu, C.; Jing, J.; Shao, H.; Jiang, Z.H.; Yan, P.; Zhu, X.-W.; Hong, W.; Zhu, A. Full-Angle Digital Predistortion of 5G Millimeter-Wave
Massive MIMO Transmitters. IEEE Trans. Microw. Theory Tech. 2019, 67, 2847–2860. [CrossRef]
8. Chung, A.; Ben Rejeb, M.; Beltagy, Y.; Darwish, A.M.; Hung, H.A.; Boumaiza, S. IQ Imbalance Compensation and Digital
Predistortion for Millimeter-Wave Transmitters Using Reduced Sampling Rate Observations. IEEE Trans. Microw. Theory Tech.
2018, 66, 3433–3442. [CrossRef]
9. Sweeney, C.; Mayeda, J.C.; Lie, D.Y.; Lopez, J. A study on Broadband Millimeter-Wave 2-Stage Power Amplifier Design in 40-nm
GaN pHEMTs. In Proceedings of the 2023 IEEE 3rd International Conference on Electronics Communications, Internet of Things
and Big Data 2023 (ICEIB), Taichung, Taiwan, 15–17 April 2023.
10. Mengozzi, M.; Gibiino, G.P.; Angelotti, A.M.; Florian, C.; Santarelli, A. GaN power amplifier digital predistortion by multi-
objective optimization for maximum RF output power. Electronics 2021, 10, 244. [CrossRef]
11. Peng, J.; He, S.; Wang, B.; Dai, Z.; Pang, J. Digital Predistortion for Power Amplifier Based on Sparse Bayesian Learning. IEEE
Trans. Circuits Syst. II Express Briefs 2016, 63, 828–832. [CrossRef]
12. Mkadem, F.; Fares, M.C.; Boumaiza, S.; Wood, J. Complexity-reduced Volterra series model for power amplifier digital predistor-
tion. Analog. Integr. Circuits Signal Process. 2014, 79, 331–343. [CrossRef]
13. Fehri, B.; Boumaiza, S. Baseband Equivalent Volterra Series for Digital Predistortion of Dual-Band Power Amplifiers. IEEE Trans.
Microw. Theory Tech. 2014, 62, 700–714. [CrossRef]
Electronics 2023, 12, 2869 19 of 19

14. Schetzen, M. Nonlinear system modeling based on the Wiener theory. Proc. IEEE 1981, 69, 1557–1573. [CrossRef]
15. Morgan, D.R.; Ma, Z.; Kim, J.; Zierdt, M.G.; Pastalan, J. A generalized memory polynomial model for digital predistortion of RF
power amplifiers. IEEE Trans. Signal Process. 2006, 54, 3852–3860. [CrossRef]
16. Ding, L.; Zhou, G.T.; Morgan, D.R.; Ma, Z.; Kenney, J.S.; Kim, J.; Giardina, C.R. Memory polynomial predistorter based on the
indirect learning architecture. In Proceedings of the Global Telecommunications Conference, 2002. GLOBECOM’02, Taipei,
Taiwan, 17–21 November 2002; IEEE: Piscataway, NJ, USA, 2002; Volume 1, pp. 967–971.
17. National Instruments. Fitting General Polynomial Vis. Available online: https://zone.ni.com/reference/en-XX/help/371361R-
01/gmath/general_polynomial_fit/#Output1 (accessed on 27 March 2023).
18. National Instruments. Memoryless DPD. Available online: https://zone.ni.com/reference/en-XX/help/374264N-01
/rfmxspecan/memoryless_dpd/ (accessed on 27 March 2023).
19. National Instruments. PXIe5831 Specifications. Available online: https://www.ni.com/pdf/manuals/377805c.pdf (accessed on
27 March 2023).
20. RF-Lambda. Available online: https://www.rflambda.com/pdf/poweramplifier/RFLUPA18G47GCK.pdf (accessed on 27 March 2023).
21. Liu, B.; Yi, X.; Yang, K.; Liang, Z.; Feng, G.; Choi, P.; Boon, C.C.; Li, C. A carrier aggregation transmitter front end for 5-GHz
WLAN 802.11 ax application in 40-nm CMOS. IEEE Trans. Microw. Theory Tech. 2019, 68, 264–276. [CrossRef]
22. Park, C.S.; Sundström, L.; Wallén, A.; Khayrallah, A. Carrier aggregation for LTE-advanced: Design challenges of terminals. IEEE
Commun. Mag. 2013, 51, 76–84. [CrossRef]
23. Guan, N.; Wu, N.; Wang, H. Digital predistortion of wideband power amplifier with single undersampling ADC. IEEE Microw.
Wirel. Compon. Lett. 2017, 27, 1016–1018. [CrossRef]
24. Guan, N.; Wu, N.; Wang, H. Model identification for digital predistortion of power amplifier with signed regressor algorithm.
IEEE Microw. Wirel. Compon. Lett. 2018, 28, 921–923. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.