Anda di halaman 1dari 9

BAI LAYOUT 1/25/12 5:22 PM Page 178



LTE-Advanced Modem Design:

Challenges and Perspectives
Dongwoon Bai, Cheolhee Park, and Jungwon Lee, Mobile Solutions Lab, Samsung Electronics
Hoang Nguyen, Jaspreet Singh, Ankit Gupta, and Zhouyue Pi, Dallas Telecommunication Lab, Samsung
Taeyoon Kim, Chaiman Lim, and Min-Goo Kim, Modem Systems Lab, Samsung Electronics
Inyup Kang, Modem Team, Samsung Electronics

ABSTRACT International Telecommunication Union (ITU)

The commercial deployment of LTE Release The evolution of LTE toward LTE-Advanced
8 is gaining significant momentum all over the makes it possible to meet IMT-Advanced
globe, and LTE is evolving to LTE-Advanced, requirements by the introduction of new features
which offers various new features to meet or such as carrier aggregation, enhanced intercell
exceed IMT-Advanced requirements. Since interference coordination (ICIC) for heteroge-
LTE-Advanced targets ambitious spectral effi- neous networks (HetNets), and enhanced multi-
ciency and peak throughput, it poses tremendous ple-antenna transmission supporting up to eight
system design challenges to operators and manu- downlink layers [1]. These new features require
facturers, especially for mobile terminals. This significant improvement of the user equipment
article discusses modem design issues related to (UE, the LTE term for handsets or mobiles),
carrier aggregation, enhanced ICIC for HetNet, and pose various design challenges. For exam-
detection of eight-layer transmission, reference ple, the need for advanced UE receivers has
signals for enhanced multi-antenna support, and been pointed out in [2] for UE to enjoy the full
HARQ buffer management. We provide an benefit of HetNet deployment.
overview of technical challenges and sketch the In this article, we focus on how this recent
perspectives for tackling them to exploit the full advancement of 3GPP standardization can affect
benefits of the LTE-Advanced system. UE receiver operation (i.e., reliable reception of
LTE-Advanced downlink transmission). More
INTRODUCTION specifically, we discuss advanced UE receiver
signal processing algorithms to address physical-
Long Term Evolution (LTE), a cellular wireless layer challenges. While most new LTE-Advanced
communication standard based on orthogonal downlink features are discussed in the following
frequency-division multiplexing (OFDM), is now sections, relaying and coordinated multipoint
under commercial deployment by many cellular (CoMP) transmission are not covered. For relay,
operators. LTE technology provides many the UE does not need to be aware of relay oper-
enhancements compared with its predecessors, ation since the relay is transparent to the UE
including high peak data rate, low latency, and [3]. Thus, UE can benefit from deployment of
spectrum flexibility. The ongoing deployment of relays without identifying them, and relaying
LTE is based on Third Generation Partnership does not impose big challenges for Release 10
Project (3GPP) Release 8. In addition to minor UE design. For CoMP, its full support has been
enhancements in LTE Release 9, such as delayed and is under active discussion for
enhanced dual layer transmission, 3GPP has Release 11.
continued its effort to improve LTE with addi- In this article, we provide a high-level
tional features. Thanks to this 3GPP effort, overview of the modem design challenges in sup-
LTE-Advanced Release 10 was recently finalized porting the Release 10 features, and suggest pos-
and frozen; that is, the base feature set for sible solutions. The main body of this article
Release 10 has been agreed. The primary goal of covers five primary issues for LTE-Advanced
Release 10 standardization is to meet the Inter- modem design: carrier aggregation, enhanced
national Mobile Telecommunications (IMT)- ICIC for HetNet, detection of eight-layer trans-
Advanced requirements for the fourth mission, reference signals for enhanced multi-
generation (4G) standards as defined by the antenna support, and hybrid automatic repeat

178 0163-6804/12/$25.00 © 2012 IEEE IEEE Communications Magazine • February 2012

BAI LAYOUT 1/25/12 5:22 PM Page 179

Filter #1 With carrier

Carrier #1 Carrier #2
ƒ LTE-Advanced
Band A supports system
bandwidth up to
Filter #1 Filter #2 100 MHz, with the
Carrier #1 Carrier #2 potential of
ƒ achieving more than
Band A 1 Gb/s throughput
(b) for downlink and

Filter #1 Filter #2
500 Mb/s
Carrier #1 Carrier #2 throughput
for uplink.
Band A Band B
Configured component carriers

Carrier #1 Carrier #2 Carrier #3 Carrier #4 Carrier #5

Activated Deactivated Deactivated Activated Deactivated

Figure 1. LTE-Advanced carrier aggregation scenarios: a) intra-band contiguous component carriers; b)

intra-band non-contiguous component carriers; c) inter-band component carriers; d) configured compo-
nent carriers with activation/deactivation.

request (HARQ) buffer management. Conclud- vation medium access control or the SCell de-
ing remarks are given in the last section. activation timer maintained by the UE. This
concept is shown in Fig. 1d. An SCell can be de-
activated, e.g., when there is not enough traffic
CARRIER AGGREGATION for the SCell. This allows a UE terminal to turn
LTE-Advanced supports aggregation of multiple off part of the transceiver chain (e.g., radio fre-
component carriers. This feature enables system quency [RF] front-end and fast Fourier trans-
deployments with large bandwidth to achieve form [FFT]) to save power. Due to this
very high data rates, allows operators to deliver activation/deactivation process, the state of
a better user experience by aggregating scattered SCells can be changed frequently. As a result,
spectrum, and supports interference manage- radio link monitoring is only supported in the
ment in heterogeneous networks via cross-carri- PCell but not in SCells to avoid complex UE
er scheduling. With carrier aggregation, behavior and additional control signaling over-
LTE-Advanced supports system bandwidth up to head. Despite this kind of effort, the increased
100 MHz, with the potential of achieving more complexity due to physical downlink control
than 1 Gb/s throughput for downlink and 500 channel (PDCCH) decoding and timing tracking
Mb/s throughput for uplink. of multiple component carriers cannot be avoid-
The component carriers can be contiguous ed. However, it does not create any complicated
within the same band, as shown in Fig. 1a, non- technical issue, but extra hardware including
contiguous in the same band, as shown in Fig. additional PDCCH decoding and timing tracking
1b, or in different bands as shown in Fig. 1c. blocks for additional component carriers and
Each of the aggregated component carriers can memory buffer to cope with a relative propaga-
have a different bandwidth. The component car- tion delay difference among component carriers
riers that a UE needs to support are determined at the UE side.
through a UE-specific configuration process and The receiver RF filter design depends on the
a dynamic activation and deactivation process. type of carrier aggregation. For carrier aggrega-
Each UE is configured with a primary compo- tion with intraband contiguous component carri-
nent carrier (also known as Primary Cell or ers, either a single RF filter or two separate RF
PCell). All the other component carriers can be filters can be used. To use a single FFT to cover
configured as secondary component carriers the contiguous component carriers, the carriers
(also known as Secondary Cells or SCells). The should be separated by integer multiples of 300
configured SCells are subject to activation/de- kHz — the least common multiple of the LTE
activation through either the activation/de-acti- channel raster (100 kHz) and LTE subcarrier

IEEE Communications Magazine • February 2012 179

BAI LAYOUT 1/25/12 5:22 PM Page 180

spacing (15 kHz). In that case, the contiguous restrict public access if it has a closed subscriber
In SIC, ordering component carriers can be separated in the digi- group (CSG). Thus, it is more difficult to man-
tal baseband. The advantage of using a single age cross-tier interference between a macrocell
does matter, and RF filter is low hardware complexity due to a and a femtocell than between a macrocell and a
the strongest single RF chain for carrier aggregation. Howev- picocell.
interference can be er, it requires an analog-to-digital converter and The use of carrier aggregation is a possible
RF filter with larger bandwidth. Moreover, due solution to manage cross-tier interference. For
cancelled first. If the to activation and deactivation of component car- example, control signaling of a macrocell is
interference is not riers, there is a retuning issue when a single RF assigned at one carrier, while control signaling of a
is used for these component carriers. When the pico/femtocell is assigned at the other carrier. The
strong enough to be status of an SCell is changed from activated to partition of control channels eliminates the possi-
decoded, it can be deactivated, the UE can save power by retuning bility of cross-tier interference. However, depend-
ignored. its RF from an aggregated bandwidth including ing on the availability of spectrum for operators,
the SCell to a smaller bandwidth excluding the carrier aggregation may not be employed.
SCell. However, during this retuning procedure, In the cochannel deployment scenario with-
some packets may be dropped due to RF transi- out carrier aggregation, resources should be par-
tion time for retuning. In LTE, the impact of RF titioned in a more sophisticated manner to
retuning to PDCCH is exacerbated because the manage cross-tier interference because two het-
PDCCH is located at the beginning of each sub- erogeneous cells share the same bandwidth. To
frame — a region likely to fall into the transition minimize this cross-tier interference, an almost
due to RF retuning. Loss of a PDCCH leads to blank subframe (ABS) is introduced for
loss of the entire physical downlink shared chan- enhanced ICIC, where most subcarriers remain
nel (PDSCH) subframe in the absence of control unused, and thus its interference to other cell is
information required for decoding, which causes limited. Even in ABS, some reference signals,
increased packet errors. RF retuning is closely like the common reference signal (CRS), and
related to the measurement of the deactivated some essential control signals, such as the physi-
SCell since the UE needs to retune its RF when- cal broadcast channel (PBCH), the primary syn-
ever the deactivated SCell measurement is chronization signal (PSS), and the secondary
required. Thus, LTE-Advanced should balance synchronization signal (SSS), need to be trans-
the packet loss due to RF retuning and the mea- mitted to support legacy UE of Release 8/9.
surement of deactivated SCells to achieve opti- Now, the necessity for advanced UE receivers
mal network performance. and their modem design challenge come into
For carrier aggregation with non-contiguous play. Since CRS, synchronization signals, and
component carriers (in either the same band or PBCH are still transmitted in ABS, they
different bands), separate RF filters and FFTs inevitably collide with signals from other cells.
are needed. Moreover, harmonics and intermod- Even with recent standardization effort, this type
ulation products from an uplink interband carri- of collision is unavoidable. An advanced UE
er aggregation transmitter can desensitize the receiver with interference cancellation tech-
receive band or have considerable impact on niques will still benefit from HetNet cochannel
other radio technology in the handset. Thus, ele- deployments, while legacy UE of Release 8/9
ments of the RF like the transmit/receive filter can maintain its connection to the macrocell
with proper attenuation should be carefully con- thanks to interference management through
sidered. Moreover, to support multiband carrier ABS. The rationale behind this interference can-
aggregation, new UE modems need to include cellation technique is that if the interference is
additional components such as band switch, strong, it can be decoded and cancelled one by
diplexer, and duplexer. The additional insertion one from the strongest interference; this is
loss due to these components needs to be han- known as successive interference cancellation
dled properly by using a higher-gain power (SIC). In SIC, ordering does matter, and the
amplifier or a better duplexer, and relaxing the strongest interference can be cancelled first. If
RF requirement or implementation margin. the interference is not strong enough to be
decoded, it can be ignored. However, imperfect
interference cancellation can degrade the quality
ENHANCED ICIC FOR HETNET of measurements used for radio link monitoring,
A HetNet consists of low-power picocells and radio resource management, and channel quality
femtocells in addition to high-power macrocells indicator (CQI) feedback.
[2]. In contrast to a traditional homogeneous To illustrate the effectiveness of interference
deployment of macrocells where the data rate of cancellation, let us consider the cell search pro-
end users at the cell edge can suffer, the deploy- cedure under enhanced ICIC as an example. If
ment of picocells and femtocells can improve overlapping or adjacent cells are frame synchro-
end-user experience connected to these cells. nized, even with ABS, synchronization signals
However, this benefit comes at the cost of addi- (i.e., PSS and SSS) of these cells will collide. For
tional intercell interference between heteroge- a homogeneous deployment, this may not be a
neous cells. While a HetNet can be deployed in problem since the UE just needs to attach to the
LTE Release 8/9, this cross-tier interference is cell providing higher downlink power (i.e., the
effectively handled by enhanced ICIC, a recently cell with higher correlation of synchronization
developed interference coordination feature in signals). However, in a HetNet deployment, the
LTE-Advanced Release 10. Note that picocells UE may want to attach to the cell providing
are deployed by operators, while femtocells are lower downlink power. This could happen if the
mostly deployed by end users, possibly in an access to the cell with higher power is restricted
unplanned manner. Moreover, a femtocell can (e.g., a CSG femtocell).

180 IEEE Communications Magazine • February 2012

BAI LAYOUT 1/25/12 5:22 PM Page 181

Figure 2a illustrates the cell search procedure

with SIC. In this procedure, the UE indentifies Detect the Yes
Start cell strongest All cells List cell
cells one by one starting from the cell with the search found? search results
highest power. However, PSS and SSS of the
detected cell are cancelled before the next cell is No
detected. This SIC on synchronization signals
can be applied until all cells are detected. In Fig. Cancel detected
2b, we consider a two-cell deployment scenario
where the target cell power is 10 dB less than
the interfering cell power. It is clearly shown
that the use of SIC is essential to reduce the 100
detection error in the HetNet cell search proce-
dure with enhanced ICIC. In general, the
advanced UE receiver needs the capability of
interference mitigation to enjoy the full benefit
of HetNet deployment with the enhanced ICIC

Probability of detection error

feature of LTE-Advanced. 10-1

The LTE-Advanced system should fulfill the
ITU requirements of the downlink peak spectral
efficiency of 30 b/s/Hz. The peak spectral effi-
ciency is the highest achievable data rate per
overall cell bandwidth assuming error-free con- PSS+SSS miss, with SIC
PSS+SSS miss, without SIC
ditions when all available radio resources for the 10-3
corresponding link direction are assigned to a -10 -5 0 5 10
single UE unit. While two- or four-layer trans- SNR (dB)
mission would be more prevalent even for the (b)
LTE-Advanced system, the required downlink Figure 2. a) Cell search procedure for ehanced ICIC with SIC; b) cell search
peak spectral efficiency can only be attained performance.
using high-order antenna configurations (i.e., 8 ×
8). The main challenge of high-order multiple-
input multiple-output (MIMO) is computational Thus, with a careful choice of K, it is possible to
complexity. Maximum likelihood (ML) detection achieve performance close to ML with much
is optimal in the sense that it minimizes error lower computational complexity. Nevertheless,
probability when the distributions of all data are the computational complexity of K-best is still
equally likely. However, due to its high complex- relatively high to be implemented in a real hand-
ity in 8 × 8 MIMO systems with high modulation set with high-order antenna configurations.
order, a direct implementation of ML detection The challenge in LTE-Advanced system
might not be a viable option for such MIMO sys- design is therefore to devise a MIMO detector
tems. Instead of ML detection, linear MIMO that achieves performance close to ML, while
detection algorithms having lower complexity still incurring computational cost within the
than optimal ML detection, such as zero-forcing reach of current hardware. Existing near ML
(ZF) or minimum mean square error (MMSE), MIMO detectors such as K-best only take us
can be applied with interference cancellation halfway in that direction. Moreover, the soft bit
techniques. However, both still have greatly infe- value calculation of the MIMO detector [6], and
rior performance to ML detection. the use of iterative detection and decoding [4]
Thus, several near ML algorithms such as can be studied to further enhance their perfor-
sphere decoding (SD) [4] and K-best [5] have mance. In short, more research effort is required
been introduced. Those algorithms are unceas- to realize the gains of MIMO systems with the
ingly being improved to achieve performance hardware constraints.
close to ML with lower complexity. In particular,
the classical K-best detector attempts to output a
list of size K consisting of the most likely values REFERENCE SIGNALS FOR ENHANCED
of the transmitted symbols based on observa-
tions, and is used for hard detection of symbols
[5]. As K increases, the computational complexi- In LTE/LTE-Advanced, the evolved NodeB
ty increases while the performance improves. (eNodeB, the LTE term for a base station) trans-
Figure 3 shows the block error rate (BLER) mits certain predefined reference signal
performance and computational complexity com- sequences along with the data. These are
parisons of K-best with different K values and employed by the UE to estimate the downlink
MMSE detection algorithms for an 8 × 8 MIMO channel for the twin purposes of feeding back
system. It can be seen that the performance channel state information (CSI) to the eNodeB
improvement of K-best becomes marginal as the and equalization of the downlink channel in the
value of K exceeds 256 for 16-quadrature ampli- process of data demodulation. In LTE/LTE-
tude modulation (QAM), while the computa- Advanced, there are four non-overlapping CRS
tional complexity of K-best grows linearly with K. patterns (corresponding to four transmit antenna

IEEE Communications Magazine • February 2012 181

BAI LAYOUT 1/25/12 5:22 PM Page 182

1 106
K-Best (K = 64)
105 K-Best (K = 256)
K-Best (K = 1024)

Number of operations
0.1 104


0.01 102

K-Best (K = 64) 101
K-Best (K = 256)
K-Best (K = 1024)
0.001 100
15 17 19 21 23 25 Real multiplication Real addition Real division
SNR (dB)
(a) (b)

Figure 3. BLER performance and computational complexity comparison (8 × 8 MIMO with two code words, 16-QAM, code rate 0.8): a)
BLER performance comparison; b) computational complexity comparison.

ports), as shown in Fig. 4. Since the CRS is used Typically, algorithm design for CSI feedback
for data demodulation, the density of the refer- involves the computation of an effective signal-
ence signals is high. to-noise ratio (SNR) metric. The effective SNR
One of the requirements for LTE-Advanced is is a physical layer abstraction meant to convert a
that it should support up to eight-layer transmis- set of different SNRs, experienced during coded
sion, which implies that there need to be at least transmission over a frequency selective fading
eight transmit antenna ports. Toward this, a key channel, into an equivalent SNR that would
change in the reference signal design philosophy result in a similar BLER when transmitting over
from LTE Release 8 is the separation of the the static additive white Gaussian noise
demodulation reference signals (DM-RS) from (AWGN) channel [7]. Using the computed effec-
the channel state information reference signals tive SNR, and precomputed reference BLER vs.
(CSI-RS) in Release 10. This potentially saves sig- SNR curves for the AWGN channel, the UE can
nificant reference signal overhead, as it allows the feed back the largest CQI that meets the target
densely populated DM-RS to be UE-specific. In a BLER requirement. With ideal channel state
particular resource block (RB), the eNodeB needs information at the UE, the preceding abstraction
to transmit DM-RS only for the layers that are is known to work well. However, using standard
being transmitted to the scheduled UE, irrespec- channel estimation at the CSI-RS RE locations,
tive of the number of transmit antenna ports. On we observe that the effective SNR prediction can
the other hand, CSI-RS continues to be transmit- be significantly erroneous. In Fig. 5a, we consid-
ted from each of the antenna ports at the eNodeB er 50 different channel realizations and plot the
to enable full CSI feedback from the UE. following three effective SNRs:
In light of the new reference signal designs • Predicted effective SNR with ideal CSI
for LTE-Advanced, significant new challenges • Predicted effective SNR with CSI-RS-based
emerge from a modem design perspective, as channel estimation
discussed next. • The actual effective SNR
We can observe that with ideal CSI, the predic-
CSI-RS tion matches the actual effective SNR closely,
In order to provide for eight CSI-RS patterns, while with estimated CSI, there can be signifi-
the density of CSI-RS in LTE-Advanced is sig- cant errors. This can result in inaccurate CQI
nificantly less than that of CRS. Specifically, feedback, potentially causing the downlink
there is only one CSI-RS resource element (RE) BLER to overshoot the mandated target BLER.
per RB per antenna port (Fig. 4). Furthermore, Next, we consider the impact of CSI-RS chan-
unlike CRS, which is transmitted in every sub- nel estimation on throughput performance. This
frame (every 1 ms), CSI-RS is expected to be entails not only CQI feedback, but also, and per-
transmitted only once every five or ten sub- haps more crucially from a throughput perspec-
frames. The sparse nature of CSI-RS poses a tive, the spatial preprocessing (i.e., PMI/RI)
major technical hurdle for UE modem develop- feedback, because channel estimation errors can
ment in LTE-Advanced. In particular, advanced result in suboptimal PMI/RI feedback, leading
algorithms are needed for CSI-RS-based channel to throughput degradation. For a 2 × 2 MIMO
estimation and computation of the CSI feedback system, Fig. 5b depicts the throughput as a func-
parameters such as CQI (i.e., modulation and tion of the SNR. We observe that channel esti-
coding rate) because existing methods may incur mation errors result in about 2 dB performance
throughput degradation and/or result in a failure loss in the low SNR regime.
to meet the target BLER requirements. Our Our results demonstrate the need to investi-
results in Fig. 5, discussed next, provide a gate advanced algorithms for CSI-RS channel
glimpse of some of these issues. estimation and CSI feedback. Possible strategies

182 IEEE Communications Magazine • February 2012

BAI LAYOUT 1/25/12 5:22 PM Page 183

The effective SNR is

0,1,4,6 0,1,4,6
a physical layer
2,3,5,7 2,3,5,7 abstraction meant to
convert a set of dif-
1 3 0 0 1 1 2 0
ferent SNRs, experi-
4 3 enced during coded
transmission over a
frequency selective
0 2 1 0,1,4,6 0 3 1 0,1,4,6 fading channel, into
2,3,5,7 2,3,5,7 an equivalent SNR
that would result in
a similar BLER when
1 3 0 2 3 1 2 0 transmitting over the
static additive white
6 7
Gaussian noise
0,1,4,6 0,1,4,6 (AWGN) channel.

0 2 1 2,3,5,7 0 3 1 2,3,5,7


CRS ports 0,1 CRS ports 2,3 CSI-RS ports 0-7 DM-RS ports 0-7
(multiplexed in pairs of 2)

Figure 4. Reference signals: locations and density of CRS, CSI-RS, and DM-RS.

to improve the CSI-RS channel estimation per- ing information because both data and DM-RS
formance could include exploitation of the time- use the same precoder or beamforming matrix
frequency correlations (across different RBs within the same RBs. The precoder itself is not
and/or different subframes), for example, using a specified in the specification, and its usage
2-D MMSE filter. This is accompanied by a depends on eNodeB vendors.
caveat that such approaches hinge critically on In general, DM-RS channel estimation
the estimation of the channel’s power delay pro- involves an interpolation using the reference sig-
file (PDP) and the Doppler shift, which pose sig- nals. It is well known that a 2-D MMSE filter is
nificant technical challenges of their own. an optimal linear filter minimizing the mean
Potential benefits of exploiting these correlations squared error of channel estimation [9]. The 2-D
are indicated by the results in Fig. 5b, wherein MMSE filter uses correlation information both
we have depicted the throughput gains obtained in the frequency and time domains. Since the 2-
using a 2-D MMSE interpolator, with known D filter is computationally more complex than
PDP and Doppler shift. two similar 1-D filters, 1-D filters are frequently
adopted and applied separately in the time and
DM-RS frequency domains. However, the lower com-
LTE-Advanced DM-RS is expected to be used plexity of 1-D filters comes at the cost of a high-
mainly for multilayer transmission supporting up er error floor, as shown in Fig. 5c. Lower
to eight DM-RS antenna ports. DM-RS mapping complexity methods of 2-D MMSE filters need
to REs is illustrated in Fig. 4. DM-RS is multi- to be identified to make them viable for imple-
plexed by a hybrid scheme of code-division and mentation in UE modems. In addition, it is
frequency-division multiplexing. For code-divi- shown that the DM-RS channel estimation per-
sion multiplexing, the time-domain orthogonal formance can be improved by considering the
cover code (OCC) or Walsh sequence is used properties of time-varying channels.
[8]. At each UE receiver antenna, channel esti- With the advent of advanced LTE-Advanced
mates associated with each transmission layer transmission techniques, some aspects of DM-
(or antenna port) are obtained by despreading RS-based channel estimation merit investigation.
the received reference signals with the known First, the baseline of DM-RS channel estimation
reference signal sequence and OCC. This orthog- is performed per RB. This single RB-based
onal despreading with the scrambled OCC is channel estimation may restrict the availability
valid when the channel is constant over refer- and performance of channel estimation methods.
ence signals (i.e., under static or slowly time- For example, frequency domain correlation
varying channels). Regarding precoding, UE can information required for MMSE-based interpo-
demodulate data in the PDSCH without precod- lation cannot be directly obtained on a per-RB

IEEE Communications Magazine • February 2012 183

BAI LAYOUT 1/25/12 5:22 PM Page 184

the channel estimation performance through RB

7 bundling [8]. Second, multi-user interference
Actual effective SNR
Predicted effective SNR with ideal CSI besides intercell interference also needs to be
Predicted effective SNR based on CSI-RS
6 considered when multi-user MIMO (MU-
MIMO) schemes are used. When MU-MIMO is
configured, the eNodeB may try to select pre-
coding matrices that minimize the multi-user
interference among co-scheduled UE. In prac-
Effective SNR

4 tice, however, the precoded DM-RS at co-sched-

uled UE may not be perfectly orthogonal to
3 each other. Due to non-orthogonal precoding
and reference signal sequences of MU-MIMO
transmission, interference cancellation schemes
are required to obtain reasonable channel esti-
mation performance. In practice, SIC can be
adopted for interference cancellation, where
interference is eliminated by repeating estima-
0 tion and subtraction schemes.
5 10 15 20 25 30 35 40 45 50
Channel realization
2.5 The data channel uses a turbo code for forward
CSI-RS-based estimated channel
Ideal CSI
CSI-RS-based estimated channel
error correction (FEC) and employs a stop-and-
(known PDP and Doppler) go protocol, hybrid automatic repeat request
2 (HARQ), where a codeword can be punctured
and transmitted in multiple attempts. Which of
the bits are punctured in a given attempt is indi-
Throughput (Mb/s)

1.5 cated by the redundancy version index. For each

HARQ attempt that fails decoding, the UE
sends a negative acknowledgement and waits for
1 the next retransmission attempt.
HARQ brings many benefits, including
throughput maximization, latency control with
time-interlaced HARQ processes, and fine con-
trol of system resource usage. However, HARQ
is effective only if the UE has the memory to
store the soft bits (i.e., soft channel bits) after
-4 -2 0 2 4 6 8 10 12 decoding fails. This is because the probability of
SNR (dB) successful decoding of a single HARQ transmis-
(b) sion can be very low, especially when the base
station overestimates the channel strength or
2-D MMSE w/ considering time-varying channel performs aggressive puncturing, where soft com-
2-D MMSE w/o considering time-varying channel
1-D linear interpolation bining of multiple HARQ attempts is needed to
ensure successful decoding.
Storage requirement is perhaps the biggest
disadvantage of HARQ. Compared to LTE
Releases 8 and 9, LTE-Advanced requires larger
Mean square error (dB)

HARQ storage due to the higher throughput.
Thus, it becomes crucial to reduce the storage
requirement by managing the HARQ buffer effi-
ciently. Table 1 lists the required buffer sizes for
different UE categories, of which the last three
are new additions for LTE-Advanced. Note that
the UE capability is classified into several cate-
gories in LTE-Advanced [10]. User equipment
from category 5, 6, or 7 must be able to store
about 3.7 million soft bits. For UE category 8,
0 5 10 15 20 25 30 the number is increased tenfold. Such storage
SNR (dB) requirements pose a big challenge for handset
(c) modem designers. Suppose that it takes four
logic gates to store one bit, and each soft bit is
Figure 5. CSI-RS and DM-RS channel estimation. Impact of CSI-RS channel represented by an 8-bit integer. Then, UE cate-
estimation errors on the a) accuracy of effective SNR prediction b) achieved gory 5 would require a gate count of about 117
throughput. c) DM-RS channel estimation performance with MMSE and lin- million just for the HARQ data buffering. Thus,
ear interpolation. without efficient techniques for HARQ buffer
management, such storage requirements would
mean a very large die size for on-chip memory.
basis, since the number of reference signals is Fortunately, it is not necessary to store the
very limited. Nevertheless, UE may use the pre- soft bits at the original resolution. A soft bit is
coding granularity in multiple RBs, and improve typically in the form of log-likelihood ratio, which

184 IEEE Communications Magazine • February 2012

BAI LAYOUT 1/25/12 5:22 PM Page 185

provides a high-level overview of UE modem

UE category Total number of soft bits
design challenges in supporting LTE-Advanced HARQ brings many
Release 10 features and suggests possible solu-
Category 1 250,368 tions. Various challenges in the following areas benefits, including
are identified, and their implications and solu- throughput maxi-
Category 2 1,237,248 tions to LTE-Advanced modem design covered: mization, latency
carrier aggregation, enhanced ICIC for HetNets,
Category 3 1,237,248 detection of eight-layer transmission, reference control with time-
signals for enhanced multi-antenna support, and interlaced HARQ
Category 4 1,827,072 HARQ buffer management. By overcoming the
challenges in UE modem design, LTE-Advanced processes, and fine
Category 5 3,667,200 will deliver on its promise of significant enhance- control of system
ments in the end-user experience. resource usage.
Category 6 3,654,144
Category 7 3,654,144 The authors would like to thank Dr. Hairuo effective only if the
Zhuang, Dr. Chun Kin Au Yeung, Dr. Jung
Hyun Bae, and Dr. Andrey Rog for providing UE has the memory
Category 8 35,982,720
simulation results. The authors would also like to store the soft bits
Table 1. Total number of soft bits supported by to thank Dr. Sudhir Ramakrishna for coordinat-
LTE-Advanced UE categories [10]. after decoding fails.
ing collective efforts in Dallas.

can be represented by 8 bits accurately. When [1] S. Parkvall, A. Furuskär, and E. Dahlman, “Evolution of
decoding fails, the soft bits can be compressed to LTE Toward IMT-Advanced,” IEEE Commun. Mag., vol.
a significantly lower resolution before sending to 49, no. 2, Feb, 2011, pp. 84–91.
storage. In some cases a 1-bit compressed resolu- [2] A. Damnjanovic et al., “A Survey on 3GPP Heteroge-
neous Networks,” IEEE Wireless Commun., vol. 18, no.
tion is able to maintain a throughput loss of less 3, June 2011, pp. 10–21.
than 1 dB in SNR. However, in high-data-rate [3] A. Ghosh et al., “LTE-Advanced: Next-Generation Wire-
scenarios, such low resolution causes significant less Broadband Technology,” IEEE Wireless Commun.,
throughput losses, especially when the effective vol. 17, no. 3, June 2010, pp. 10–22.
[4] B. M. Hochwald and S. ten Brink, “Achieving Near-
FEC code rate is too high for a given SNR. Capacity on A Multiple-Antenna Channel,” IEEE Trans.
To improve the throughput performance in Commun., vol.51, no.3, March 2003, pp. 389–99.
challenging scenarios, the compressed resolution [5] Z. Guo and P. Nilsson, “Algorithm and Implementation
needs to be higher than 1 bit. In particular, frac- of the K-Best Sphere Decoding for MIMO Detection,”
IEEE JSAC, vol. 24, no. 3, Mar. 2006, pp. 491–503.
tional rates can be achieved using vector quanti- [6] J. Lee, J.-W. Choi, and H. Lou, “MIMO Maximum Likeli-
zation techniques [11]. Because soft bits hood Soft Demodulation Based On Dimension Reduc-
belonging to the same modulation symbol are tion,” Proc. IEEE GLOBECOM, 2010.
more strongly correlated than bits that do not, it [7] J. Zhuang et al., “IEEE 802.16m Evaluation Methodolo-
gy Document,” IEEE 802.16m-08/004r4, Nov. 2008.
is beneficial to group together soft bits that [8] 3GPP, ‘’Evolved Universal Terrestrial Radio Access (E-
come from the same modulation symbol when UTRA, Physical Channels and Modulation (Release 10),”
vectorizing the log-likelihood ratio sequence to TS 36.211, V10.2.0, June 2011.
be compressed. [9] P. Hoeher, S. Kaiser, and P. Robertson, “Pilot-Symbol-
Aided Channel Estimation in Time and Frequency,”
In addition to compression, HARQ buffer Proc. IEEE GLOBECOM, Commun. Theory Mini-Conf.,
memory needs to be further managed via non- 1997, pp. 90–96.
compressive techniques in case of carrier aggrega- [10] 3GPP, “Evolved Universal Terrestrial Radio Access (E-
tion. For example, the required buffer size for UE UTRA); User Equipment (UE) Radio Access Capabilities
(Release 10),” TS 36.306 V10.2.0, June 2011.
category 5 is 3,667,200 soft bits divided equally [11] R. M. Gray, “Quantization,” IEEE Trans. Info. Theory,
among all configured downlink component carri- vol. 44, no. 6, Oct. 1998, pp. 2325–83.
ers. The eNodeB, however, assumes a soft buffer
size of 3,667,200 for rate matching per component BIOGRAPHIES
carrier, regardless of the number of component D ONGWOON B AI ( is a staff
carriers. While the number of soft bits received engineer at Samsung Mobile Solutions Laboratory, San
can exceed the required storage space, the UE can Diego, California. He received his B.S. degree in electrical
choose to store more soft bits than the specifica- engineering from Seoul National University in 1998 and his
M.S. degree in electrical engineering from KAIST in 2000.
tion requires. Because not all configured compo- He received his S.M. degree in applied mathematics and
nent carriers are fully utilized at all times, the UE Ph.D. degree in engineering science from Harvard Universi-
can dynamically divide HARQ storage space ty in 2006 and 2009, respectively. From 2009 to 2011, he
among component carriers by assigning less/no worked at AppliedMicro, Sunnyvale, California.
storage to underutilized/idle component carriers CHEOLHEE PARK ( is a staff engi-
and more to loaded component carriers. However, neer with the Mobile Solutions Laboratory, Samsung R&D
a mechanism should be in place to rebalance the Center, San Diego, California, where he works on the LTE
memory usage when an underutilized/idle compo- modem design. Before he joined Samsung, he was with
the Satellite and Mobile Terminal Division of Hyundai Elec-
nent carrier becomes busy. tronics Industry Co. from 1997 to 1999 and with the Wire-
less Network Research Center of Korea Electronics
Technology Institute from 1999 to 2004. He received his
CONCLUSION Ph.D. degree in electrical and computer engineering at the
University of Texas at Austin in 2010.
LTE Advanced is a flexible wireless broadband
technology that promises significant enhance- JUNGWON LEE ( is a director at
ments in the end-user experience. This article Samsung Mobile Solutions Laboratory, San Diego, Califor-

IEEE Communications Magazine • February 2012 185

BAI LAYOUT 1/25/12 5:22 PM Page 186

nia. He received his Ph.D. degree in electrical engineering joining Samsung, he worked at Nokia Research Center in
from Stanford University in 2005. From 2003 to 2010, he Dallas and San Diego on 3G standardization and modem
was with Marvell Semiconductor Inc. as a principal engi- development. He has published more than 20 research
neer and staff manager, where he led a wireless systems papers and more than 80 patents and patent applications.
and research group consisting of more than 15 engineers He received his B.E. degree in sutomation from Tsinghua
with Ph.D. degrees. Since 2010 he has been with Samsung University (with honor), his M.S. degree in electrical engi-
U.S. R&D Center. He has co-authored more than 40 papers neering from Ohio State University, and his M.B.A. degree
as well as standards contributions for various standards from Cornell University (with distinction).
bodies, and holds over 30 U.S. patents. His main research
interests lie in wireless and wireline communication theory. T AEYOON K IM ( has been a
senior systems engineer with Samsung Electronics since
HOANG NGUYEN ( is a staff engi- October 2009, where he works on the system design of
neer at Samsung Telecommunication America, Dallas, modem chipsets and 3GPP standardization work. Prior to
Texas, where he works on 3G/4G wireless systems. Prior to joining Samsung, he was with Cellular Product Group at
joining Samsung, he was a research engineer with Nokia Freescale Semiconductor, Austin, Texas from 2006 to
Inc. and then a member of technical staff with Lincoln Lab- 2009. He received his Ph.D. degree in electrical and com-
oratory at Massachusetts Institute of Technology (MIT). He puter engineering from the University of Texas at Austin
received his B.S. degree (summa cum laude) in 1999 from in 2006.
the University of Missouri-Columbia, and M.S. and Ph.D.
degrees in 2002 and 2003 from the University of Califor- CHAIMAN LIM ( is a principal engineer
nia–Davis, all in electrical engineering. He is a member of in the Samsung Modem Team, Suwon, Korea, where he is
the Phi Kappa Phi and Tau Beta Pi honor societies. He working on modem chipset development. He has been
holds a Missouri registered EIT license, five issued patents, with Samsung Electronics since 1998 where he is primarily
and has many patents pending. He has authored over 20 involved in the system design of 3G (WCDMA/HSPA) and
journal and conference papers. He is a Graduate Research 4G (LTE) modems for terminals. His current research inter-
Fellow of the U.S. National Science Foundation. ests include advanced MIMO and LTE-Advanced.

J ASPREET S INGH ( is a senior MIN-GOO KIM ( joined Samsung

research engineer at the Samsung R&D Center, Dallas, Electronics in 1993 and is currently the head of groups in
Texas, where he is working on advanced algorithms for the area of mobile systems and RF technologies. He has
LTE-A terminal modem development. He received his been deeply involved in the research and development of
B.Tech. degree in electrical engineering from the Indian 3G (WCDMA/HSPA) technologies for the applications for a
Institute of Technology, Delhi, in 2004, and M.S. and Ph.D. terminal. His current research interests include LTE and LTE
degrees in electrical and computer engineering from the evolution. He holds a Ph.D. in electronics engineering from
University of California, Santa Barbara, in 2005 and 2009, Seoul National University (SNU), Korea.
respectively. He was a summer intern at the Swiss Federal
Institute of Technology, Lausanne, in 2003, and at Qual- INYUP KANG ( received B.S. and
comm Flarion Technologies, New Jersey, in 2008. M.S. degrees in electronics engineering from Seoul Nation-
al University in 1985 and 1987, respectively, and a Ph.D.
A NKIT G UPTA ( received his degree in electrical engineering from UCLA in 1996. From
B.Tech. degree in 2003 from the Indian Institute of Tech- 1988 to 1991 he was at Korea Telecom Research Center.
nology, Delhi, and M.A. and Ph.D. degrees in 2006 and From 1996 to 2009 he was at Qualcomm, San Diego,
2009, respectively, from Princeton University, New Jersey, where he led generations of cellular modem chipsets as VP
all in electrical engineering. He is a senior research engi- of technology. Since 2010 he has been working at Sam-
neer with Samsung Telecommunications America. sung Electronics as EVP in charge of R&D and commercial-
ization of cellular baseband and RF chipset and software
ZHOUYUE PI ( is a director with Sam- for 2G, 3G, and 4G. His main research interests are in the
sung R&D Center, Dallas, Texas, where he leads next gener- fundamental and theoretical aspects of cellular wireless
ation standardization and development efforts. Before communications.

186 IEEE Communications Magazine • February 2012