This article has been accepted for inclusion in a future issue of this magazine.

Content is final as presented, with the exception of pagination.

ACCEPTED FROM OPEN CALL

Massive MIMO in Sub-6 GHz and mmWave:

Physical, Practical, and Use-Case Differences
Emil Björnson, Liesbet Van der Perre, Stefano Buzzi, and Erik G. Larsson

Abstract in sub-6 GHz and mmWave bands. In this article,

we provide a comparative overview, highlighting
The use of base stations (BSs) and access points three main differences:
(APs) with a large number of antennas, called • The propagation channels build on the same
Massive MIMO (multiple-input multiple-output), physics, but basic phenomena such as diffrac-
is a key technology for increasing the capac- tion, attenuation, and Fresnel zones are sub-
ity of 5G networks and beyond. While origi- stantially different.
nally conceived for conventional sub-6 GHz • The hardware implementation architecture
frequencies, Massive MIMO (mMIMO) is also changes with the increasing carrier frequency.
ideal for frequency bands in the range 30–300 More antennas can be integrated into a given
GHz, known as millimeter wave (mmWave). area, but the insertion losses, intrinsic pow-
Despite conceptual similarities, the way in er-overhead in radio-frequency (RF) generation,
which mMIMO can be exploited in these bands and amplification result in diminishing gains.
is radically different, due to their specific prop- • The signal processing algorithms depend on
agation behaviors and hardware characteristics. propagation and hardware. Channel estima-
This article reviews these differences and their tion is resource-demanding at sub-6 GHz, while
implications, while dispelling common misun- beamforming is straightforward. Conversely,
derstandings. Building on this foundation, we mmWave channel estimation and beamforming
suggest appropriate signal processing schemes are theoretically simpler since there are fewer
and use cases to efficiently exploit mMIMO in propagation paths, but become challenging if
both frequency bands. hybrid beamforming is used.
In the remainder of this article, we elaborate on
Introduction these differences, including that they manifest how
mMIMO uses arrays with many antennas at the sub-6 GHz and mmWave bands are to be exploit-
BS to provide vast signal amplification by beam- ed to target different use-cases in 5G and beyond.
forming and high spatial resolution to multiplex
many simultaneous users. Although small-scale Difference I: The Propagation Channel
MIMO technology has been around for decades, An understanding of electromagnetic propagation
the practical gains have been modest due to the is crucial when considering mMIMO systems and
small number of antennas which seldom give suf- frequencies up to mmWave bands. The channels
ficient spatial resolution to support many spatially behave fundamentally different from what we are
multiplexed streams. mMIMO has been demon- used to in cellular networks, which exposes weak-
strated to achieve an order-of-magnitude higher nesses in the channel modeling simplifications
spectral efficiency in real life, with practical acqui- conventionally made.
sition of channel state information (CSI) [1]. 3GPP
is steadily increasing the maximum number of
antennas in LTE and since 64 antennas are sup-
Sub-6 GHz:
ported in Release 15, mMIMO has become an Favorable Propagation and Spatial Correlation
integral component of 5G. Radio channels below 6 GHz have been widely
Another key approach to increase the capac- studied for single-antenna and small-scale MIMO
ity of future wireless networks is the operation in systems. The propagation depends on path-loss
mmWave bands. There are many GHz of unused and shadowing, called large-scale fading, and
spectrum above 30 GHz, which can be used as multi-path propagation, resulting in small-scale
a complement to the current sub-6 GHz bands. fading. In recent years, measurement campaigns
The path-loss and blockage phenomena are more have been carried out to characterize sub-6
severe in mmWave bands, but can be (partially) GHz mMIMO channels [2, 3]. For example, the
overcome by keeping the same physical size of real-time testbed at Lund University, shown in
the antenna array as on lower frequencies, which is Figure 1a, has substantially contributed to the
achieved by mMIMO. There are, however, funda- understanding of both mMIMO propagation
mental differences between how mMIMO technol- phenomena and hardware implementation. Fig-
ogy can be designed, implemented, and exploited ure 1b shows an alternative distributed mMIMO

Emil Björnson is with Linköping University; Liesbet Van der Perre is with Katholieke Universiteit Leuven and with Lund University;
Digital Object Identifier: Stefano Buzzi is with the University of Cassino and Lazio Meridionale and with the Consorzio Nazionale Interuniversitario per le
10.1109/MWC.2018.1800140 Telecomunicazioni (CNIT); Erik G. Larsson is with Linköping University.

2 1536-1284/19/$25.00 © 2019 IEEE IEEE Wireless Communications • Accepted for Publication

 This article has been accepted for inclusion in a future issue of this magazine. Content is final as presented, with the exception of pagination.

deployment. mMIMO measurements show that

the UEs’ channels become closer to orthogonal
with an increasing number of antennas, referred
to as favorable propagation [4]. Differently from
small-scale MIMO, large-scale fading can poten-
tially vary significantly between the antennas in
mMIMO. This occurs, for example, when part
of a physically large array is more shadowed
than the rest [5, Sec. 7.3] or when using cylindri-
cal arrays where the antennas point in different
directions.
Many researchers consider i.i.d. Rayleigh fad-
ing channels in their assessment of mMIMO. This
approach is analytically tractable, provides insight-
ful rate expressions, and leads to channel harden-
ing, where the impact of small-scale fading reduces
as more antennas are added. In contrast, correlat-
ed fading channels are more complicated to ana-
lyze [5]. However, in practice, not every channel
is well-modeled as having i.i.d. channel coefficients.
Some UEs will have strong line-of-sight (LoS) com- (a) Co-located mMIMO testbed at Lund University.
ponents and many UEs will feature spatially cor-
related small-scale fading. These characteristics
must be modeled to capture how spatial correla-
tion leads to more (less) interference between UEs
that have similar (different) spatial correlation char-
acteristics [5]. The channel hardening effect is also
weaker under spatial correlation.
Although the beamforming becomes more
directive as the number of antennas M increas-
es, it has no effect on how frequently we need to
re-estimate the channel under mobility. To show
this, consider a UE in LoS that moves a fraction
m ≤ 1/8 of the wavelength. The mth channel
coefficient is phase-shifted by ej2pfm, where fm 
[–m, m] depends on the direction of movement.
If the beamforming is fixed at the original UE
location, the beamforming gain will reduce from
M to |S M m=1e
j2pf m| 2/M ≥ |S M cos(2pf )| 2/M
m=1 m
≥ Mcos2(2pm) ≥ M/2. Hence, in the worst case,
the beamforming gain is reduced by 3 dB, inde-
pendently of M. This might seem counterintuitive,
i.e., the beamwidth becoming narrower as M
increases, but is explained by the fact that we need
to be further away before the emitted signal takes
the form of a beam. In conclusion, we never need
to estimate the channel more often than it takes to
(b) Distributed mMIMO testbed at KU Leuven.
move 1/8th of a wavelength and usually much less FIGURE 1. Photos of two operational real-time mMIMO testbeds, both built
frequently (the chain of inequalities is very conser- using hardware components from National Instruments and using a 20
vative). MHz bandwidth: a) the co-located testbed at Lund University supports
100 BS antennas around a carrier frequency of 3.7 GHz; b) the testbed at
mmWave: Blessing and Curse of KU Leuven supports 64 BS antennas designed for the 2.4–2.62 GHz and
Attenuation and Directivity 3.4–3.6 GHz bands and supports distributed operation. (Photos courtesy of
U. Lund — EIT and KU Leuven — ESAT.)
The measuring and modeling of mmWave
channels have received considerable attention,
leading to a solid understanding of how these area is the same as on lower frequencies. The
channels differ from sub-6 GHz channels [6] latter gives the flexibility to change the directiv-
and extensions of the 3GPP channel models ity of the array by beamforming, which is highly
to support carrier frequencies from 0.5 to 100 desirable in mobile communications. The feasibil-
GHz (see 3GPP TR 38.901). We first consider ity of communicating at a high rate in LoS, ben-
large-scale fading. Recalling the Friis transmis- efiting from the wide available bandwidth, also
sion equation, the smaller wavelength l directly over long distances has been exploited using
increases the path-loss proportionally to l–2. This high-gain directional antennas.
is due to the implicit assumption of fixed-gain The total beamforming gain of a communica-
antennas whose effective area is proportional to tion link is the product of the beamforming gains at
l2. Hence, it can be overcome by using fixed-ar- the transmitter and receiver. Instead of deploying
ea antennas, which become increasingly direc- a huge array at one side of the link, the same total
tional with a gain proportional to l–2, or using an beamforming gain can be achieved by deploying
array of fixed-gain antennas whose total effective substantially smaller arrays at both sides. For exam-

IEEE Wireless Communications • Accepted for Publication 3

 This article has been accepted for inclusion in a future issue of this magazine. Content is final as presented, with the exception of pagination.

In mmWave systems, a ple, instead of having 1000 BS antennas to serve baseband. However, neither the digital processing
multi-GHz bandwidth single-antenna UEs, we can have 100 BS antennas nor the data converters are a complexity hurdle,
requires ADCs which, and 10 antennas per UE. This also opens the door although those are the stages where hybrid beam-
for physical reasons, to explore systems with massive arrays at both forming primarily induces simplifications. It is the
sides [7]. high-speed interconnect that is a bottleneck in the
consume an order of The Fresnel zone defines the region around the realization of integrated systems.
magnitude more power LoS path that should be non-obstructed to avoid Next, we concisely discuss the key hardware
than their counterparts severe signal losses. Its radius, at a point located at sub-systems in mMIMO processing: the digital
for sub-6 GHz systems, distances d1 and d2 from the two ends of the link, baseband and data converters, the RF and analog
considering similar respectively, is given by sub-systems, and the interconnection of the many
antenna signals.
linearity requirements. d1 d2 / (d1 + d2 ).
When equipping UEs
with antenna arrays,
At 38 GHz, the Fresnel zone has ≈ 0.5 m radi-
us for a communication distance of 100 m. For
DSP and Data Converters:
the ADC power con- shorter distances and higher frequencies, it goes Lean Processing Suits the System
sumption will impact down to the cm-range. Hence, the Fresnel zone We distinguish three main parts in mMIMO DSP:
can be obstructed by small objects, leading to • The outer modem applying error-correction
hybrid beamforming abrupt channel variations even when the trans- coding on each data stream individually. Its
trade-offs. mitter and receiver are fixed. In mobile access, complexity is not impacted by mMIMO.
the signal strength will fluctuate rapidly as the • Central processing performing decoding and
obstruction changes. In combination with highly transmit beamforming. Operations on large
directive antennas, this calls for antenna arrays matrices can be implemented efficiently in
deliberately capturing reflections, or fast electron- hardware when exploiting the nearly orthogo-
ic beam-switching to reflected paths, if they are nal user channels [9].
available. • Antenna signal processing by which we mean
At mmWave frequencies, many objects the DSP performed on signals connecting data
behave as full blockers, including humans [8], converters to the central processing. Their over-
and there is less diffraction. Specific frequencies all complexity scales with the number of full
suffer from absorption by gases with colliding res- (RF and analog) front-end chains from anten-
onance frequencies, such as 60 GHz for oxygen. nas to digital baseband. This part may be domi-
Losses > 40 dB have been measured through a nant in terms of operations/second, but can be
window, which is substantially higher than for implemented at low resolution [5, Sec. 6].
sub-6 GHz waves. The outdoor-to-indoor cov- Taking advantage of scaled CMOS technolo-
erage is therefore rather limited in mmWave gy and system-level opportunities, efficient DSP
bands. Outdoors, significant losses through foli- implementations are feasible in both sub-6 GHz
age have also been observed [6]. Rain will cause and mmWave bands.
higher attenuation with increasing frequencies, Data converters are a potential bottleneck in
but the impact on the link budget is rather small. hardware complexity in multi-antenna process-
A consequence of these unfavorable propaga- ing. Low-power architectures have rendered this
tion effects is that the link budget is worse in objection obsolete for BSs. Analog-to-digital con-
mmWave bands than at sub-6 GHz, even if we verter (ADC) cores achieve figures-of-merit in
let the physical size of the BS antenna array be terms of energy consumption per conversion step
the same in both bands. (cs) in the order of 30 fJ/(fs2ENOB), where ENOB
Small-scale fading will also be considerably stands for “effective number of bits.” For each bit
different with only one-bounce reflected paths reduction in resolution, the ADC power is basical-
actually contributing. The reflected paths may ly halved. For low resolutions even 10 fJ/cs has
allow communication in case the LoS is blocked. been reported [10].
Since small-scale fading changes substantially ADCs in 4G systems (sub-6 GHz) require a
when moving 1/8th wavelength, 10 times faster resolution ENOB > 10, owing to the dynamic
channel variations occur at 30 GHz than at 3 range requirement imposed by the combination
GHz when moving at the same speed, which of OFDM, MIMO, and high-order constellations.
calls for 10 times more frequent channel esti- mMIMO in realistic conditions is expected to
mation. This might be less of an issue in prac- work well with ENOB = 5 [5, Sec. 6.4.1]. Hence,
tice; the coverage area of mmWave BSs is rather the ADC power consumption in a 128-antenna
small, thus only low-mobility UEs will likely con- BS can be lower than in a conventional 8-antenna
nect to them. system. The actual power consumption of an inte-
grated ADC may be a factor of 2–4 higher than
Difference II: Hardware Implementation in theory [11] to account for voltage regulators,
In mMIMO, an evident concern is the implemen- input buffering, and calibration. Still, an individual
tation complexity of the digital baseband and ana- converter may consume less than 1 mW. A few
log/RF hardware. Technology scaling has fueled hundred of them is hence negligible in the total
impressive progress in wireless communication power budget of a BS.
systems and is essential to process many antenna In mmWave systems, a multi-GHz bandwidth
signals. requires ADCs which, for physical reasons, con-
The flexibility offered by full digital beamform- sume an order of magnitude more power than
ing leads to the highest theoretically achievable their counterparts for sub-6 GHz systems, consid-
performance, while hybrid analog-digital beam- ering similar linearity requirements. When equip-
forming schemes are explored to enable hard- ping UEs with antenna arrays, the ADC power
ware reuse over antenna paths, by having a mixed consumption will impact hybrid beamforming
RF signal chain from the antennas to the digital trade-offs.

4 IEEE Wireless Communications • Accepted for Publication

 This article has been accepted for inclusion in a future issue of this magazine. Content is final as presented, with the exception of pagination.

Generating, Phase Shifting,

and Amplifying Rf Signals: Divide and Conquer?
The synthesis of RF frequencies is challenged by
strict constraints on phase-noise and error vector
magnitude (EVM). These requirements are tough-
er to meet at mmWave compared to sub-6 GHz.
The oscillator efficiency is highly influenced by the
ratio of the operating frequency over the channel
spacing and the Q-factor of the resonator. For the
former, one may assume channel spacing to go
up with operating frequency. The Q-factor of the
resonator typically drops at mmWave, resulting in
a lower oscillator efficiency [11].
Hybrid beamforming implemented with analog
phase-shifting on the RF signals maximizes hard-
ware reuse. For mmWave systems, the phase-shift-
ers need to be able to settle fast to sustain
communication when the LoS path is disrupted.
The realization of precise phase-shifting is dif-
ficult at high frequencies and may incur consider-
able power overhead. Hence, implementing the
phase-shifting in the analog baseband may be pre- FIGURE 2. mmWave module hosting a 4-antenna transceiver IC co-integrated
ferred. with patch antennas (two patches for each of the four antenna paths). The
The power amplifier (PA) constitutes the most size is 5.4  9.2 mm. (Photo courtesy of imec.)
power-hungry component in RF transceivers.
Linear PAs are required for the 5G broadband
transmission schemes. mMIMO systems benefit lenges; interconnects are the main bottleneck to
from reduced output power requirements, both exploit the high bandwidth through the integra-
for the entire array and per antenna [5, Sec. 5.2]. tion of many small antennas.
Their combined complexity (cost) and power will
decrease with a growing number of antennas, with Difference III: Signal Processing Algorithms
diminishing returns. The major differences in channel propagation
A sub-6 GHz PA operating at 6 dB back-off can and hardware implementation have fundamental
achieve a power added efficiency (PAE) of 18 per- implications on the algorithms needed for channel
cent [12]. mmWave PAs need to rely on power estimation, beamforming, and resource allocation.
combining, which introduces extra losses. More-
over, the lower gain at these frequencies calls for Opportunities for Efficient Channel Estimation
higher DC drive currents [11]. CMOS PAs achieve The number of channel coefficients grows linearly
PAE< 10 percent at 6 dB back-off. with the number of antennas at the BS and UE. To
Operating PAs closer to saturation has been have an approximate idea of the computational
suggested for mMIMO systems in order to increase burden, consider a system with 200 BS antennas
power-efficiency. However, this may infringe on and 20 spatially multiplexed single-antenna UEs.
the specification for EVM and out-of-band radiation Consider OFDM with 1024 subcarriers and chan-
as coherent combining of the non-linear distortion nels that are constant over 12 subcarriers. There
will occur in scenarios with a few dominant beams are 3.4 · 105 complex scalar coefficients, which
[13], as expected in mmWave. amounts to 6.8 · 106 estimates/second if a chan-
nel coherence time of 50 ms is assumed. These
Interconnect is the Main Implementation Challenge numbers increase if there are more antennas,
mMIMO systems process a large number of more subcarriers, and/or shorter coherence time.
antenna signals. Connecting these signals con- At sub-6 GHz, there is generally multi-path
stitutes the main hardware implementation chal- propagation caused by a multitude of scattering
lenge. For sub-6 GHz systems, in order to bring clusters. The channel coefficients are correlated
all individual signals to the DSP level, a balanced across antennas, but this can only be utilized to
approach with partly distributed processing can marginally improve the estimation quality, at the
circumvent the bottleneck [9]. cost of substantially higher complexity [5, Sec. 3].
At mmWave, the connections to the antennas Nevertheless, the estimation can be conveniently
become extremely lossy since micro-strip lines implemented/parallelized in hardware [9] and the
behave as antennas, giving losses of several dB/ estimation overhead is small when operating in
cm at 60 GHz for different integration materi- time-division-duplex (TDD) mode and exploiting
als [14]. Matching of components is challenging channel reciprocity to only send uplink pilots [4].
[11]. Systems will only benefit from more anten- At mmWave, the channel can potentially be
nas if these can be integrated in a very compact parameterized (considering a phase-synchro-
way, urging a co-design of chips, antennas, and nized array with a known angular array response)
package, as illustrated in Fig. 2 for the transceiver because it consists of a (potential) LoS path and
described in [15]. few one-bounce reflections. Instead of estimating
Unfortunately, hybrid beamforming does not the individual channel coefficients, a few angular
relax the requirement of connecting mmWave channel coefficients can be estimated to acquire
signals to antennas. Oscillator distribution at the entire channel, leading to greatly reduced com-
mmWave frequencies also faces severe chal- plexity. When a single data-stream is to be sent, it

IEEE Wireless Communications • Accepted for Publication 5

 This article has been accepted for inclusion in a future issue of this magazine. Content is final as presented, with the exception of pagination.

suffices to estimate the dominant angle-of-arrival/ CSI acquisition, which grows with the number of
departure, but also reflections can be taken into antennas since the beams become narrower.
account. However, if hybrid beamforming is used, While TDD operation is preferable at sub-6 GHz
the phase-shifters create a very directional “vision” mMIMO, in mmWave bands frequency-division-du-
and only channel components that fall into the plex (FDD) may be equally good since the chan-
analog beams can be estimated. To discover new nel-describing angular parameters are reciprocal
UEs, track channel variations, or keep the connec- over a wide bandwidth.
tion when the LoS path is blocked, beam-sweeping
is needed (i.e., the channel must be estimated in
many different directions to identify the preferable
Choosing Between Analog,
ones). This procedure increases the overhead from Hybrid, and Digital Beamforming
Current hardware can realize full digital beam-
forming at sub-6 GHz, while hybrid analog/digital
beamforming is a potential design-simplification at
mmWave. With analog transmit beamforming, a
phase-shifted version of the same signal is trans-
mitted from all antennas. This leads to a signal
beam directed in a particular angular direction
(see Fig. 3a). If multiple UEs are multiplexed, one
Digital set of analog phase-shifters (connected to a sep-
baseband arate input from the baseband) is needed per
UE (see Fig. 3b). This is hybrid beamforming in
a nutshell. The number of UEs cannot be larger
than the number of baseband-inputs, but digital
Phase-shifters precoding can be used to assign a mix of the UEs’
(a) Analog beamforming: Only one beam is created for the entire frequency band, which signal to each input.
is sufficient for LoS beamforming. In contrast, full digital beamforming can send
any signal from any antenna. This flexibility can be
exploited at sub-6 GHz frequencies to deliver high
beamforming gain in rich multi-path environments,
as illustrated in Fig. 3c. The digital flexibility is evi-
dent in multi-user scenarios, where the antennas
should transmit a superposition of many beams
per UE, different beams per subcarriers (due to fre-
Digital quency-selective fading), and multiplex many UEs
baseband on the same time-frequency resource slot. Ana-
log beamforming cannot adapt the beam to multi-
path propagation and frequency-selective fading,
while hybrid beamforming has only a limited ability
Phase-shifters to do that since the phase-shifters create a set of
(b) Hybrid beamforming: A few beams are created by the analog phase-shifters and the fixed beams that the digital precoding needs to be
digital baseband can create superpositions of these beams to adapt to multi-path and based on.
frequency-selective fading, which gives a limited flexibility for handling NLoS scenarios.
mmWave systems have lower user multiplexing
capability if implemented with hybrid beamform-
ing, since the number of UEs is limited by the num-
ber of sets of phase-shifters.
However, even analog beamforming (Fig. 3a)
suffices for single-user communications over wide
bandwidths. To illustrate this fact, we consider a
LoS channel with five reflections. The center fre-
quency is 60 GHz and we use the method in [5,
Digital
baseband Sec. 7.3.2] to compute the array response for dif-
ferent frequencies. We consider 32  32, 64  64,
and 128  128 planar arrays with half-wavelength-
spaced antennas.
The maximum beamforming gain is equal to the
number of antennas and is achievable by digital
beamforming. Figure 4 shows the percentage of
the maximum beamforming gain that is obtained
by analog beamforming at different frequencies
around the center frequency. It starts at 90 percent
since no analog beamforming matches the array
(c) Digital beamforming: Full flexibility to create a superposition of any number of beams response when there are reflections. Nevertheless,
and adapt the beams to multi-path and frequency-selective fading. if the bandwidth is 400 MHz, 80–90 percent of
the maximum beamforming gain can be achieved
FIGURE 3. The hardware implementation of beamforming determines the in the entire band by analog beamforming. If the
flexibility in handling difficult propagation scenarios, which in turn has impli- bandwidth continues to grow, the beamforming
cations on the use cases. The system is viewed from above and different gain drops since the beamforming is optimized
colors represent different signals: a) analog beamforming; b) hybrid beam- for the center frequency. This is known as the
forming; and c) digital beamforming. beam-squinting effect. The gain loss is particularly

6 IEEE Wireless Communications • Accepted for Publication

 This article has been accepted for inclusion in a future issue of this magazine. Content is final as presented, with the exception of pagination.

severe for larger arrays, since the beams are nar-

100
rower. With the 32  32 array, more than 75 per-

Percentage of maximum beamforming gain (%)

cent of the maximum gain can be achieved over a 90
2 GHz bandwidth, while the gain drops quickly for
the 128  128 array. 80

Network Deployment for Good Coverage 70

Several differences between sub-6 GHz and
mmWave arise in network planning and resource 60
32 ×32
allocation. At sub-6 GHz, BSs ensure both out-
50 64 ×64
door and indoor coverage, and support to 128×128
high-mobility UEs. Although the BS positioning is 40
important, it is not as crucial as the adoption of
interference-management procedures, such as 30
advanced digital beamforming techniques that
deal with inter-cell interference and pilot contami- 20
nation [5, Sec. 4].
In contrast, at mmWave frequencies, given the 10
blockage effects, very careful deployment planning 0
is needed to provide coverage to an intended area. 59 59.5 60 60.5 61
Interference is less important, but ensuring wide-ar- Frequency [GHz]
ea coverage, without coverage holes, may require
a large number of BSs. FIGURE 4. The fraction of the maximum beamforming gain that is achieved
at different frequencies when analog beamforming is used. The phases
Use-Cases: are selected to maximize the inner product with the array response at the
center frequency. There is a LoS path with azimuth angle p/4 and elevation
Different Solutions for Different Cases angle – p/4 to the array, measured from the boresight. There are also five
Although the data traffic increases by 30–40 per- reflections, with the azimuth angles p/6, p/3, p/4, p/4, p/12 and the corre-
cent annually, contemporary macro-cells only sponding elevation angles – p/5, – p/5, – p/6, – p/12, and – p/6. The total
need to serve one or a few UEs at any given gain of the LoS equals the gain of all the reflections.
time instant. The reason is that the networks
have been gradually densified. In traffic-intense
areas, the inter-BS distance is in the order of 100 conditions, since both the desired and interfering
m, rendering further densification questionable signals become larger, the beamforming gain does
from a practical and cost perspective. Hence, improve for the cell-edge UEs by only increasing
the number of simultaneous UEs is likely to the desired signals. Hence, it can be utilized to
grow rapidly in the future. The new use cases provide uniformly high quality-of-service through-
of ultra reliable low latency communication out the cell. With 40 MHz bandwidth and 3 b/s/
(URLLC) and massive machine-type communi- Hz, data rates of 120 Mb/s can be achieved per
cations (mMTC) to support diverse Internet-of- UE. By multiplexing 20 UEs, the cell throughput
Things (IoT) applications are two drivers toward becomes 2.4 Gb/s.
this change. A world with sensors everywhere, mMIMO at sub-6 GHz offers the same support
autonomous cars, drones and social robots, and for user mobility as other technologies operating
augmented-reality applications will require a net- in that band, and high-mobility support has been
work infrastructure that supports 100 times high- demonstrated in field-trials [3].
er capacity than today. Since the purpose of using mmWave bands
The key use-cases and the propagation scenar- is to have 10–100 times more bandwidth than
ios are summarized in Table 1. One cannot sep- at sub-6 GHz, the link budget will be reduced by
arate these aspects since a technology can be a 10-20 dB (assuming the same output power and
perfect fit for a use-case in one scenario, but infea- effective antenna area at the BS). When combined
sible in another scenario; for example, mMIMO with the fact that outdoor-to-indoor propagation
in mmWave bands can provide unprecedented is rather limited and the signals are easily blocked,
data rates in LoS scenarios, but is less suited for a huge number of BSs, relays, and/or reflective
outdoor-to-indoor communications. To deliver all surfaces would be needed to guarantee wide-area
the necessary services, we need to evolve the net- coverage. The mmWave band is, however, attrac-
works in two respects: tive for providing fixed wireless access over large
• Improve the macro-cell BSs to handle many areas, since the BSs and UEs can then be deployed
simultaneous UEs. to guarantee LoS-like conditions.
• Deploy short-range BSs that offload traffic in Hotspots: Auditoriums, cafés, airports, and sta-
hotspots. diums are examples of hotspots where the data
Macro-cells: mMIMO at sub-6 GHz is ideal traffic is very high in a physically small area. To off-
for delivering higher throughput in macro-cells load the macro-cells, WiFi is most frequently used
than in legacy networks. As noted in Table 1, the in these places, but WiFi supports neither mobility
cell-edge and outdoor-to-indoor coverage are nor high user loads. These issues can be resolved
improved by the beamforming gain: the received by using mMIMO at sub-6 GHz (an array need
useful signal power grows proportionally to the not be larger than a television screen), but since
number of antennas, whereas the (average) inter- LoS propagation dominates in hotspots, mmWave
ference power at other locations remains the same mMIMO is a more suitable solution. In hotspots, a
due to non-coherent combination. While network decent signal-to-noise ratio (SNR) can be achieved
densification does not improve average cell-edge over a huge bandwidth, thanks to the short prop-

IEEE Wireless Communications • Accepted for Publication 7

 This article has been accepted for inclusion in a future issue of this magazine. Content is final as presented, with the exception of pagination.

Feasibility and suitability of mMIMO in different uses cases.

Use case mMIMO in sub-6 GHz mMIMO in mmWave

High data rates in most propagation scenarios (e.g., ~100 Mb/s/user Huge data rates (e.g., ~10 Gb/s/user using several GHz of
Broadband access
using 40 MHz of bandwidth), with uniformly good quality-of-service. bandwidth) in some propagation scenarios (see below).

Beamforming gain gives power-savings and better coverage than Not fit for low data rate applications, which will incur significant
IoT, mMTC
legacy networks. power overhead.

URLLC Channel hardening improves reliability over legacy networks. Difficult since propagation is unreliable due to blockage.

Mobility support Same great support as in legacy networks. Very challenging, but theoretically possible.

Narrow beamforming is possible; with 100 antennas, 20 dB Beamforming gain is possibly higher than at sub-6 GHz, since more
High throughput fixed link
beamforming gain is achievable; only array size limits the gain. antennas fit into a given area, but the gain per antenna is smaller.

Spatial multiplexing of tens of UEs is feasible and has been Same capability as at sub-6 GHz in theory, but practically limited if
High user density
demonstrated in field-trials. hybrid implementation is used.

Feasibility and suitability of mMIMO in different propagation scenarios.

Propagation scenario mMIMO in sub-6 GHz mMIMO in mmWave

Outdoor-to-outdoor, indoor-to- High data rates and reliability (see above) in both LoS and NLoS Huge data rates (see above) in LoS hotspots, but unreliable due to
indoor communication scenarios. blockage phenomena.

Outdoor-to-indoor
High data rates and reliability (see above). Limited due to higher propagation losses.
communication

Can multiplex many links, even in NLoS, but relatively modest data Great for LoS links, particularly for fixed antenna deployments, but
Backhaul/fronthaul links
rates per link. less suitable for NLoS links.

Mainly interference-limited in cellular networks, due to high SNR Mainly noise-limited in indoor scenarios, due to huge bandwidth and
Operational regime
from beamforming gains and substantial inter-user interference. limited inter-cell interference, but can be interference-limited outdoors.

TABLE 1. Feasibility and suitability of mMIMO at sub-6 GHz and mmWave for different use cases and propagation scenarios.

agation distances, leading to extreme throughput. dred samples per coherence block when operating
For example, with 1 GHz of bandwidth, a spectral at sub-6 GHz, giving room to orthogonal resourc-
efficiency of 1 b/s/Hz is sufficient to achieve 1 es for channel estimation to a few hundred UEs.
Gb/s per data stream. With more spectrum and/or This number is inversely proportional to the carrier
higher spectral efficiency, 10 Gb/s is within reach. frequency [4, Sec. 2], leading to an order-of-magni-
This is a key use-case for mmWave technology. tude of fewer samples in mmWave bands.
Spatial multiplexing of UEs can be implemented In an environment without significant mobility,
using hybrid beamforming, as illustrated in Fig. 3b, very large numbers of UEs may be multiplexed [4,
if the UEs are in LoS. Since the channels evolve Sec. 6.1]. Consider, for example, a festival taking
20 times faster when going from, say, 3 GHz to place in Central Park, Manhattan. This large park is
60 GHz carrier frequency, mmWave hotspots can surrounded by skyscrapers, where BS antennas can
easily support pedestrian movement, while higher be mounted to provide LoS conditions. Eventually,
speeds are more challenging. only measurements can determine the channel
coherence, but assume for the sake of argument
What if Extra Hardware Came at no Cost? a coherence block of 100 ms by 400 kHz when
Suppose the hardware and signal processing operating at 3 GHz. The coherence time reduc-
come for free and work perfectly, how large an es to 5 ms when operating at 60 GHz.
array could eventually be useful? Figure 5 shows the downlink sum rate when
In an environment without significant mobili- operating at these frequencies, as a function of the
ty, very large numbers of users may be spatially number of antennas and UEs, and assuming that
multiplexed. In [4, Sec. 6.1], one sub-6 GHz case fully digital reciprocity-based beamforming is used
study establishes the feasibility of providing (fixed) in all cases. The sum rate grows monotonically with
wireless broadband service to 3,000 homes, using the number of antennas, as expected. The highest
a BS with 3,200 antennas (which at 2 GHz requires values on the curves are 1.38 Tb/s at 3 GHz (with
an array of around 4  4 m). By jointly increasing 50 MHz bandwidth) and 1.44 Tb/s at 60 GHz (with
the number of antennas and UEs, the total radi- 1 GHz bandwidth), which are nearly the same. The
ated power per BS and rate per UE can be kept huge difference is that the peak values are achieved
constant. by multiplexing 14,000 or 870 UEs, respectively.
The number of UEs that can be spatially mul- This corresponds to allocating 35 percent and 44
tiplexed per BS is determined by the number of percent of the coherence blocks to uplink pilots,
samples per channel coherence time-frequency respectively. The mmWave setup delivers 1.66 Gb/s
block and the number of BS antennas. An outdoor per UE, while the sub-6 GHz setup only delivers 99
network that supports high mobility has a few hun- Mb/s per UE, but compensates by serving extreme-

8 IEEE Wireless Communications • Accepted for Publication

 This article has been accepted for inclusion in a future issue of this magazine. Content is final as presented, with the exception of pagination.

1400 1500

1200

1000
Sum rate [Gbit/s]

Sum rate [Gbit/s]

1000
800

600
500
400

200

0 0
10 10
×10
4 8 ×10
4 8
Nu Nu
mb 6 mb 6
ero ero
fa 4 4 fa 4
nte nte 2000
nna 2 3 4 nna 2 1500
s( 2 ×10 s( 1000
M) 1 M)
0 UEs (K) 0 500
0 Number of 0 UEs (K)
Number of

(a) 3 GHz carrier frequency and 50 MHz bandwidth. (b) 60 GHz carrier frequency and 1 GHz bandwidth.

FIGURE 5. Downlink sum rate that is achieved when operating at different carrier frequencies using TDD and digital beamforming, as
a function of the number of BS antennas. The uplink SNR to each receive-antenna is 20 dB when operating at 3 GHz with 50 MHz
bandwidth and scaled accordingly when operating at 60 GHz with 1 GHz bandwidth to keep the transmit power fixed. The down-
link transmission uses 100 times more power than the uplink pilot transmission. Closed-form rate expressions from [4, Sec. 3.3]
were used to generate the figures. To make the signal processing complexity scalable, maximum ratio transmission and channel
estimation based on uplink pilots are assumed. a) 3 GHz carrier frequency and 50 MHz bandwidth; b) 60 GHz carrier frequency
and 1 GHz bandwidth.

ly many UEs. This exposes the fundamental opera- While this article has substantiated how
tional difference; the huge bandwidth in mmWave mMIMO offers unprecedented performance to
bands allows for high per-UE rates, while the longer end users, other applications are envisioned, such
coherence time at sub-6 GHz allows for spatial mul- as the implementation of cloud-RAN through
tiplexing of more UEs. Which solution is preferable in-band wireless fronthauling [5, Sec. 7.6]. The
depends on the data traffic characteristics of the enormous amount of baseband data available in
future, but why not deploy both? mMIMO systems can be also used to sense the
The maximum number of antennas was 100,000 environment; for example, estimate the amount of
in this futuristic simulation. Assuming 3 GHz and traffic on a road, count the number of persons in a
half-wavelength-antenna-spacing, these antennas room, or guard against intrusion in protected spac-
can be deployed in array of 31 m  31 m. At 60 es. In conclusion, as far as mMIMO is concerned,
GHz, this shrinks to 1.58 m  1.58 m. Both setups the best is yet to come.
can easily be deployed at the face of a skyscrap-
er, so the size is not an issue. However, adequate
implementation strategies are needed to cope
References
[1] T. L. Marzetta, “Noncooperative Cellular Wireless with
with bottlenecks in connecting and processing the Unlimited Numbers of Base Station Antennas,” IEEE Trans.
Wireless Commun., vol. 9, no. 11, Nov. 2010, pp. 3590–
many signals. 3600.
[2] X. Gao et al., “Massive MIMO Performance Evaluation
Conclusions and the Way Ahead Based on Measured Propagation Data,” IEEE Trans. Wireless
This article has reviewed the major differences Commun., vol. 14, no. 7, July 2015, pp. 3899–3911.
[3] P. Harris et al., “Performance Characterization of a Real-Time
in mMIMO design for sub-6 GHz and mmWave Massive MIMO System with LOS Mobile Channels,” IEEE
frequencies, concerning the propagation mech- JSAC, vol. 35, no. 6, June 2017, pp. 1244–1253.
anisms, transceiver hardware, and signal pro- [4] T. L. Marzetta et al., Fundamentals of Massive MIMO, Cam-
cessing algorithms. The impact on the various bridge University Press, 2016.
[5] E. Björnson, J. Hoydis, and L. Sanguinetti, “Massive MIMO
envisioned 5G use-cases has been explained, Networks: Spectral, Energy, and Hardware Efficiency,” Foun-
showing that both bands offer attractive propo- dations and Trends R in Signal Processing, vol. 11, no. 3-4,
sitions. Computational complexity is no longer 2017, pp. 154–655.
a main bottleneck, but less considered factors, [6] T. S. Rappaport et al., “Wideband Millimeter-Wave Propaga-
tion Measurements and Channel Models for Future Wireless
such as the interconnect of signals, both for Communication System Design,” IEEE Trans. Commun., vol.
central baseband processing and at mmWave 63, no. 9, Sept. 2015, pp. 3029–3056.
to antennas, constitute potential bottlenecks. [7] S. Buzzi and C. D’Andrea, “Energy Efficiency and Asymp-
The technology is at a more advanced stage at totic Performance Evaluation of Beamforming Structures
in Doubly Massive MIMO mmWave Systems,” IEEE Trans.
sub-6 GHz, yet challenges exist in both bands. Green Commun. and Networking, vol. 2, no. 2, June 2018,
Several intriguing questions remain unanswered: pp. 385–396.
Will mmWave mMIMO be implemented with full [8] C. Gustafson and F. Tufvesson, “Characterization of 60 GHz
digital beamforming? Which mMIMO features Shadowing by Human Bodies and Simple Phantoms,” Proc.
6th European Conf. Antennas and Propagation (EUCAP),
will be actually used in 5G networks? Will the Prague, 2012, pp. 473–477.
multiplexing capabilities ever be pushed as high [9] H. Prabhu et al., “3.6 A 60pJ/b 300Mb/s 128x8 Massive
as illustrated in the Central Park example? How MIMO Precoder-Detector in 28nm FD-SOI,” Proc. IEEE Int’l.
will data-traffic patterns and applications evolve? Solid-State Circuits Conf. (ISSCC); vol. 60, San Francisco, CA,
Feb. 2017, pp. 60–61.
Whatever the answers will be, mMIMO will cer- [10] G. Van der Plas and B. Verbruggen, “A 150 MS/s 133
tainly play a paramount role in the shaping of muW 7 bit ADC in 90 nm Digital CMOS,” IEEE J. Solid-State
future wireless networks in both bands. Circuits, vol. 43, no. 12, Dec. 2008, pp. 2631–2640.

IEEE Wireless Communications • Accepted for Publication 9

 This article has been accepted for inclusion in a future issue of this magazine. Content is final as presented, with the exception of pagination.

[11] Oral conversations with experts: Dr. Bob Verbruggen, Xil- trical and Information Technology Department at Lund Uni-
inx, on ADCs; Prof. Pietro Andreani, Lund University, on versity, Sweden. She was with the nano-electronics research
frequency synthesis; Prof. Patrick Reynaert, KU Leuven, on institute imec in Belgium from 1997 till 2015, where she
power amplifiers; and Prof. Dominique Schreurs, KU Leu- took up responsibilities as senior researcher, system archi-
ven, on microwave circuits. tect, project leader and program director. She was appoint-
[12] P. Reynaert et al., “Doherty Techniques for 5G RF and ed honorary doctor at Lund University, Sweden, in 2015.
mm-Wave Power Amplifiers,” Proc. Int’l. Symposium VLSI Design, Her main research interests are in wireless communication,
Automation and Test (VLSI-DAT), Hsinchu, 2016, pp. 1–2. with a focus on physical layer and energy efficiency. She
[13] E. G. Larsson and L. Van der Perre, “Out-of-Band Radia- was the scientific leader of the EU-FP7 project MAMMOET
tion from Antenna Arrays Clarified,” IEEE Wireless Commun. on Massive MIMO. She is an author and co-author of over
Lett., vol. 7, no. 4, Aug. 2018, pp. 610–613. 350 scientific publications.
[14] S. Brebels et al., “Technologies for Integrated mm-Wave
Antenna,” The 8th European Conf. Antennas and Propaga- Stefano Buzzi ([email protected]) is a professor in the Depart-
tion (EuCAP 2014), The Hague, Apr. 2014, pp. 727–731. ment of Electrical and Information Engineering at the Universi-
[15] G. Mangraviti et al., “A 4-Antenna-Path Beamforming Trans- ty of Cassino and Southern Latium, Italy. He is a co-author of
ceiver for 60GHz Multi-Gb/s Communication in 28nm more than 150 technical papers appearing in journals and pro-
CMOS,” Proc. 2016 IEEE Int’l. Solid- State Circuits Conf. ceedings of international conferences. His most recent research
(ISSCC), San Francisco, CA, Feb. 2016, pp. 246–247. activities concern massive MIMO technologies, mmWave wire-
less systems, and cell-free and user-centric networking architec-
Biographies
Emil Björnson ([email protected]) is an associate professor
tures. He is currently an associate editor for IEEE Transactions on
Wireless Communications, and a former associate editor of IEEE
in the Department of Electrical Engineering (ISY), Linköping Uni- Communications Letters and IEEE Signal Processing Letters. He
versity, Sweden. His research interests are multi-antenna com- has also edited three special issues for IEEE Journal on Selected
munications, radio resource management, energy efficiency, and Areas in Communications.
machine learning. He is the first author of the textbooks Massive
MIMO Networks: Spectral, Energy, and Hardware Efficiency (2017) Erik G. Larsson ([email protected]) is a professor at Linköping
and Optimal Resource Allocation in Coordinated Multi-Cell Sys- University, Sweden. He co-authored Fundamentals of Massive
tems (2013). He currently serves as an associate editor for IEEE MIMO (Cambridge, 2016) and Space-Time Block Coding for
Transactions on Communications and IEEE Transactions on Green Wireless Communications (Cambridge, 2003). Recent service
Communications and Networking. He received the 2018 Marconi includes membership on the IEEE Signal Processing Society
Prize Paper Award in Wireless Communications and best paper Awards Board (2017-2019), and the IEEE Signal Processing Mag-
awards at WCSP 2017, IEEE ICC 2015, IEEE WCNC 2014, IEEE azine editorial board (2018-2020). He received the IEEE Signal
SAM 2014, IEEE CAMSAP 2011, and WCSP 2009. Processing Magazine Best Column Award twice, in 2012 and
2014; the IEEE ComSoc Stephen O. Rice Prize in Communica-
L iesbet V an der P erre ([email protected]) is tions Theory 2015; the IEEE ComSoc Leonard G. Abraham Prize
a professor in the Department of Electrical Engineering at 2017; and the IEEE ComSoc Best Tutorial Paper Award 2018.
the KU Leuven in Belgium and guest professor in the Elec- He is a Fellow of the IEEE.

10 IEEE Wireless Communications • Accepted for Publication