II. CHALLENGES AND PROSPECTS FACING 6G HYPER-CONNECTIVITY
A. Tbps for Immersive User Experience
Since the target of 5G is to support 20 gigabits-persecond (Gbps) peak data rate and 0.1 Gbps user-experienced data rate, it is not too difficult to support advanced multimedia services such as 8K augmented reality (AR) streaming, 16K virtual reality (VR) streaming, and extended reality (XR) requiring beyond-gigabit data rates. It is however not easy to support truly immersive mobile services such as digital twin or metaverse since the virtual model accurately reflecting the physical world using XR devices or high-fidelity mobile holographic displays requires a rate of up to Tbps [2]. For example, a full-view 24K VR transmission with a 64 pixels per degree (PPD) minimum resolution, a 200 Hz minimum refresh rate, and a 300:1 compression rate requires at least 6.37 Gbps. Clearly, this level of rate cannot be supported anytime anywhere in the current 5G systems since the typical user-experienced data rate of 5G is in the range of 0.1–1 Gbps.
One way to support emerging services requiring tremendous data rates is to exploit the terahertz (THz) frequency band. The THz band offers a colossal spectrum but the THz waves suffer high attenuation due to the attendant significant path loss, molecular absorption, and atmospheric attenuation. One straightforward option to overcome these problems is to densify the network, meaning that we reduce the distance between the base station (BS) and mobile device. In this ultradense scenarios, pencil beamforming via ultra-massive multiinput multi-output (UM-MIMO) plasmonic nano-antenna arrays can be useful in providing [TeX:] $$10 \text { to } 100 \times$$ improvement in data rate [6]. While the densified small cells equipped with massive multi-input multi-output (MIMO) systems are a good fit for THz communication systems, their deployment will make the notions of cell and handover obsolete. Thus, we need to design the cell-free networks from scratch by considering various technical challenges, such as beam tracking/management, user association and BS coordination, and synchronization and initial access. Other key enabling technologies to support Tbps data rates include orbital angular momentum (OAM) multiplexing [7], full-duplex, and spatial modulation-MIMO (SM-MIMO).
B. Zero Coverage-Hole Networks
Another more important goal of 5G systems is to provide advanced wireless connectivity for a wide variety of vertical applications including self-driving cars and drones, robot-aided smart factories and healthcare, and remote control for disaster scenarios. To accommodate such applications, wireless access should be available anywhere, meaning that there should be no coverage holes. Although many network operators claim nationwide coverage for 5G, there are still many coverage holes in places such as tunnels, basements, elevators, mountains, and oceans. Moreover, higher frequencies (3.5 GHz at FR1 and 28 GHz at FR2) used in 5G systems tend to be vulnerable to blockage owing to their high directivity and severe path loss. For these reasons, more often than not we experience link failure in our daily use of 5G services.
Recently, several new approaches to enhance coverage and ensure universal and seamless connectivity have been proposed. An intelligent reflecting surface (IRS), an approach that provides an artificial path in the absence of line-ofsight (LOS) links, is a viable candidate to eliminate coverage holes, particularly in dense urban areas [8]. Wavcontrolled holographic MIMO surfaces that flexibly shape electromagnetic waves using sub-wavelength metallic or dielectric scattering particle structures is another good option [9]. In remote areas such as deserts, oceans, and mountain areas, in which the deployment of terrestrial networks (TNs) is almost impossible, non-terrestrial networks (NTNs) with satellites, high-altitude platform stations (HAPSs), and unmanned aerial vehicles (UAVs) can be an attractive solution because they ensure coverage without relying on the ground infrastructure. By the deliberate combination of TNs and NTNs [10], one can dramatically improve the coverage and thus generate the network with virtually no coverage hole.
C. Pervasive Computing for Connected Intelligence
With the advent of truly immersive mobile services, such as XR and metaverse, computational overhead has become a serious concern since these services require high energy consumption, end-to-end latency, and processor and memory overhead [11]. In 5G, mobile edge computing (MEC) has been widely used to delegate computation-intensive tasks to the edge (a local server near a BS). However, MEC would not be enough to support future applications. In a VR service, for example, a mobile VR headset must provide photorealistic and 6 degrees of freedom (6-DoF) rendered images within 20 ms of motion-to-render-to-photon (M2R2P) latency. To control this overhead solely in a mobile device or offload it to a nearby BS will be by no means easy so that new paradigm to deal with the surge of computation is required. Fortunately, as the density of wireless networks increases and beyond-gigabit transmissions can be supported using mmWave and THz band transmissions, one can exploit computational resources in nearby BSs and mobile devices that are equipped with high-performance computing processors. For instance, a vehicle equipped with a high-performance battery (1000 Wh/L) and powerful graphics processing units (GPUs) can perform environment sensing, learning, and computation-oriented tasks and send the results to the destination device. Particularly, computation offloading is useful for artificial intelligence (AI) applications, such as object detection and classification, speech recognition, and metaverse control, since these tasks usually require heavy deep neural network (DNN) computation, but the inference results are often very simple.
In order to facilitate the boundless use of computation resources over the network, one should use advanced wireless distributed computing techniques. Decentralized learning obtaining a global model from localized data via peer-to-peer communication among mobile devices can be one good choice. When security matters, federated learning where the mobile devices jointly train the global model without exchanging the local dataset with the central server would be a proper choice. Yet another option to support large-scale models is split learning, in which large-scale DNNs are divided into several pieces and then trained separately on mobile devices. For the data transmitted by mobile devices to BS, federated learning transfers the entire neural network model obtained locally, while split learning transfers the intermediate output of the divided parts of the neural network model.
III. KEY ENABLING TECHNOLOGIES FOR 6G HYPER-CONNECTIVITY
In this section, we discuss key enabling technologies to achieve beyond Gbps for immersive user experience, zero coverage-hole networks, and pervasive computing for connected intelligence.
A. Distributed & Intelligence-Aware Cell-Free Massive MIMO Network
(a) Illustrations of 6G cell-free massive MIMO network (the blue lines are the reliable links between the mobile and the densely deployed small BSs and the red lines are the blocked links between the mobile and the macro cells) and (b) CDFs of user SINR in cell-free massive MIMO networks, small cell, and centralized massive MIMO networks.
In the 6G era, a network will be heavily densified and thus, the areal density of BSs will be comparable to, or even surpass, the density of active mobile devices. As the coverage of a cell gets smaller, the distance between BS and mobile will also be shorter. To make the most of the proximity between BSs and mobile devices, a network should be designed such that a group of small cells coherently serve users in a usercentric manner (see in Fig. 2(a)) [12]. In most urban scenarios, user-centric cell association will generate multiple LOS links between BSs and mobile so that the mobile can easily achieve a high degree of macro-diversity gain and spectral efficient improvement [13], [14].
To evaluate the efficacy of the cell-free massive MIMO network, we plot cumulative distribution functions (CDFs) of user signal-to-interference-and-noise ratio (SINR) in mmWave (100 GHz) and THz (1 THz) communication scenarios. In this figure, we consider the cell-free massive MIMO systems where M = 128 small cells, each of which is equipped with N = 4 antennas, cooperatively serve K = 20 mobiles using the same time-frequency resource. For comparison, we use the small cell systems where the small cells do not cooperatively serve the whole users but serve only the users in its cell coverage and the centralized massive MIMO systems where a single macro cell equipped with [TeX:] $$N_{\text {macro }}=512$$ antennas serve the users. In every scenario under test, we fix the total transmit power of the network so that the overall energy consumption does not increase in the densified network. One can see from Fig. 2(b) that the SINR variation of the cell-free massive MIMO network is much smaller, meaning that the cell-free massive MIMO network can well suppress the inter-cell interference through the cooperative beamforming of BSs. For example, in the THz systems, the cell-free massive MIMO network can provide at least 7dB SINR to more than 95% of users while the conventional small cell network guarantees only -2.5 dB. We also observe that the SINR gain of the THz band is higher than that for the mmWave band, meaning that it can effectively compensate for the huge path loss of the THz systems. To fully enjoy the benefits of cell-free massive MIMO, we need to address the following challenges.
Intelligence-aware network architecture: In order to provide the user-centric coverage through the joint beamforming of cooperating BSs, a cell control information, channel state information (CSI), and transmit data of local BSs should be shared with the central unit (CU) through the fronthaul link. Sharing such large information among all network entities will cause a significant fronthaul overhead, not to mention the tremendous computational overhead required for the channel estimation, data transmission/reception, and resource allocation [15]. A promising approach to mitigate the overhead is decentralized processing. One such example is multi-agent deep reinforcement learning (MA-DRL) where the operations like beamforming, resource allocation, and user association are performed locally and individually by the DRL agent in each small cell.
Multi-sensory data fusion for channel prediction and beamforming: In order to realize a truly zero coveragehole cell-free massive MIMO network, the network should have mobility-aware functions such as fast-fading channel estimation, mobile trajectory tracking, blockage prediction, and dynamic BS clustering and beamforming. One promising direction is to use the multi-modal sensory data (e.g., RGB depth vision information of camera, range information of LiDAR, and SLAM-based user location) along with the computer vision techniques to identify the 3D shape of nearby wireless environments (e.g., obstacle locations, LOS and non-line-of-sight (NLOS) paths, and mobility patterns). In this scheme, using a set of panoramic images obtained from the stereo camera, the DL-based object detection can extract the location (3D Cartesian coordinate (x, y, z)) of target objects in an image and then determine its class. In acquiring the location information of a mobile device from input images, a convolutional neural network (CNN) that extracts the spatially correlated feature using the convolution filter can be employed. Once the location of the mobile is identified, by measuring the angle and distance from the BS to the target object (e.g., mobile phone, vehicle, or urban air mobility (UAM)), the mmWave and THz beam can be generated without a complicated beam management process. Indeed, from the numerical evaluations based on the processing latency and power consumption of the latest Qualcomm Snapdragon 888, recently introduced AI-focused system-ona- chip (SoC), it has been shown that the image processing latency and the power consumption of deep learning (DL)- based beamforming scheme are much smaller than those of the conventional beamforming scheme in 5G NR [16]. By employing dedicated AI processors designed with a few nanoscale CMOS technology, we expect that the DNN processing latency can be further reduced down the road.
Computation-oriented end-to-end cell-free massive MIMO network: In order to support the computation-intensive and/or latency-critical applications, the future 6G network should have the computation offloading functionality [17]. Computation offloading can be realized by utilizing the local computing power at the BSs as well as cloud computing at the CU. For example, latency-sensitive computation tasks will be processed at the BSs in proximity to the user, and the computationintensive tasks can be handled via the cooperation of BSs and CU.
To ensure successful computation offloading, an integrated cell-free network design exploring both communication and computational perspectives needs to be suggested. In general, it should be guaranteed that computing operations can be delivered by reliable and low-latency communication links through a user, BSs, and CU. From a computation perspective, edge servers located at the BSs may have limited computing capabilities in terms of computing power, memory size, and energy consumption compared to the central server at the CU. In addition, since the user needs to offload computation tasks to nearby BSs, the randomness in computation latency at each BS should be taken into account to guarantee low-latency and reliable computing services.
Integrated RF and optical X-haul provisioning: In the cell-free massive MIMO network, fronthaul and backhaul overhead is a serious yet underrepresented problem. Recently, advanced fronthaul/backhaul techniques have been proposed. Replacing the current CPRI interface with the optical fiberbased fronthaul/backhaul is a simple option but will increase the CAPEX significantly. A radio stripe network-based fronthaul and backhaul where the BSs are serially connected using daisy-chain architecture to the same cables would be a cost-effective option for the densely deployed BSs. Also, an integrated access and backhaul (IAB) network where only a few BSs are connected to traditional fiber infrastructures while other BSs wirelessly relay the fronthaul/backhaul traffic would be a promising option. Yet another simple solution is free-space optical (FSO) communication where the modulated optical beams are transmitted through the free space. When the LOS link is guaranteed, FSO-based backhaul can offer several Gbps data rates with a much lower deployment cost than the optical fiber-based backhaul. When the LOS link is absent, optical IRS, a planar metasurface that controls the optical propagation environment by adjusting the phase shifts of reflecting elements can be used to offer the virtual LOS link between the BSs and CU.
Rank-sufficient THz LOS communications using cell-free massive MIMO network: In the THz band, due to the severe path loss and high directivity, the conventional MIMO techniques exploiting the rich scattering of multipath environments will not work. In the case where the LOS path is the dominant, the cell-free massive MIMO systems can be an appealing solution to address the rank-deficiency issue of THz communications. Specifically, through the joint transmission of small BSs, the cell-free massive MIMO systems can generate a virtual multipath propagation. Further, the shortened communication distance along with the extremely large number of transmit antennas incur a spherical wavefront which creates a rich non-linear pattern of phase variations and therefore, increases the channel rank.
B. Boundless and Fully Integrated Terrestrial & Non-Terrestrial Networks
In remote and rural areas where the existing terrestrial cellular network is unreachable [18], NTNs will be a useful means of providing cost-effective and high-speed internet services. Depending on the flying altitude and coverage area, NTN devices can be classified as UAV, HAPS, and satellite. Satellite stations are further divided into geostationary orbit (GEO), medium earth orbit (MEO), and low earth orbit (LEO) based on their orbital altitude.
Satellite communication systems with different altitudes: Round-trip delay and single satellite coverage as a fraction of the earth’s surface.
In Fig. 3, we plot the round-trip delay (RTD) and single satellite coverage on the earth’s surface for various altitude levels. When a satellite revolves in orbit close to the surface of the earth, a faster response time can be guaranteed, but the satellite’s decreases owing to its proximity to the ground. For example, only three equally spaced GEO satellites are needed to can provide near-global coverage because each GEO satellite covers almost 35% of the earth’s area, but RTD in this case exceeds 0.5 s. Conversely, the RTD of the LEO satellite is less than 0.1 s, but the coverage will be just 0.2% at an altitude of 550 km. One can infer from this discussion that a large number of LEO satellites, often called LEO megaconstellation, would be a good option to guarantee the latency in the order of a few tens of milliseconds and global coverage.
We note that there exists a trade-off between the ground observer’s minimum elevation angle and link quality (see Fig. 3). If we increase the minimum elevation angle, so is the LOS probability of the satellite link; however, the beam coverage and visible time of the satellite will be reduced. To balance this trade-off, the ground station’s minimum elevation angle should be carefully determined in practice. Recently, SpaceX decided to change the minimum elevation angle of Starlink LEO satellites from [TeX:] $$40^{\circ} \text { to } 25^{\circ}$$ (granted by the FCC in April 2021) to improve coverage with a small number of LEO satellites [19].
In analyzing the NTN communication of satellite and aerial devices, areal capacity [TeX:] $$\left(\mathrm{bits} / \mathrm{s} / \mathrm{m}^2\right)$$ is no longer useful; therefore, a new metric called volumetric traffic capacity [TeX:] $$\left(\text {bits} / \mathrm{s} / \mathrm{m}^3\right)$$ that measures the quality of user experience in 3D space is needed. To obtain a sufficient volumetric traffic capacity to ensure universal global connectivity for 6G systems, the following issues must be addressed.
Mobility-aware channel modeling and tracking in 3D space: Since the NTN terminals are located at a high altitude and are moving fast with respect to the ground stations, LOS propagation paths are guaranteed, but the wireless signals suffer significant Doppler shift. The first step in addressing the distortions caused by high mobility and a large coverage area is an accurate 3D channel model that accounts for the Doppler shift and propagation delay. In addition, dynamic channel prediction based on the real-time location tracking of NTN terminals is of great importance to achieve agile and effective NTN management.
Boundless multi-layered TN/NTN architecture: To provide sufficient and continuous coverage everywhere, the large LEO satellite constellations should be combined with GEO/MEO using the coverage-aware protocol and NTN-backhaul connection. In fact, a boundless multi-layered architecture fully integrating TNs and NTNs is essential to guarantee the reliable connection in the ocean and desert, as well as high-speed trains and airplanes [20]. UAV and HAPS play a key role in improving the connection quality because they can be used as relays between the LEO/MEO/GEO satellite and ground stations. In addition, an integrated network protocol that includes multi-layer routing, network mobility management, and resource allocation should be introduced.
Seamless inter-beam/satellite/orbit handover: Owing to the high-speed movement of the LEO satellites, the maximum visible time of the satellites at the ground station is at most 10–20 m, causing a frequent handover even for a fixed ground station. To mitigate this problem, LEO satellites should be connected to nearby satellites in the same or adjacent orbits through laser or radio frequency (RF) links. This so-called inter-satellite link (ISL) connectivity supports relatively low end-to-end latency even for the long distance communication scenario. Candidate techniques to establish an efficient ISL connectivity include spatial-mode diversity and multiplexing, wavelength division multiplexing, and adaptive beam control using pointing/acquisition/tracking (PAT) process. Another challenge of NTNs is that the cell size varies according to the elevation angle of the NTN terminals; therefore, conventional handover techniques might not be suitable. To address this challenge, a forecast-based handover decision on the basis of satellite position estimation is required.
3D cell planning and fast beam management: Since the cell size varies according to the altitude and elevation angle of the satellite, we need to introduce 3D cell planning based on the orbits of satellites. A satellite with a higher altitude and lower elevation angle tends to form a larger cell size owing to its wider beam footprint on the ground. Specifically, the GEO and LEO satellites have beam widths of up to 3,500 km and 1,000 km at an altitude of approximately 600 km, respectively. In addition, 3D cell types can be classified into earth-moving and earth-fixed cells with respect to the earth’s surface, depending on whether the satellite beams can be fixed or steered. The earth-fixed cell should be designed to be highly robust to satellite beam pointing errors since even a very small error makes the satellite beam footprint placed tens of kilometers away from the initially designated area on the earth; whereas, the earth-moving cell may require more frequent handovers because of the dynamics of beam movement on the ground. As for the cell planning in multi-layered TN/NTN architecture, a truncated octahedron-shaped cell planning distinguished from the hexagon-shaped 2D cell layout is desired when considering the volumetric quotient in 3D space.
Intelligent spectrum management and NTN system optimization: To achieve seamless and broad coverage in integrated terrestrial and non-terrestrial networks, powerful spectrum management scheme is needed. To this end, intelligent and dynamic cooperation between primary and cognitive users is preferred over the semi-static spectrum sharing. Note that because the S-band (i.e., 2170–2200 MHz) is adjacent to the 3G/4G LTE band (i.e., 2110–2170 MHz), the large Doppler shift of the high-speed satellites causes severe interference in the adjacent band, deteriorating the service quality of 3G/4G TN services. A possible option to mitigate the interference is the on-board pre-compensation of the Doppler shift at the center of the beam on the ground while the residual Doppler shift can be corrected at each receiver side. To further improve spectral efficiency, we should have an appropriate frequency reuse pattern among multiple beams, available channels, and antenna gains.
C. Over-the-air Distributed Computing for Artificial Intelligence Applications
Over-the-air distributed computing scenario where a VR headset assigns linear computation tasks to four edges with uncoded/coded offloading.
A future trend in dealing with the relentless growth in computing-intensive applications is pervasive computing, in which mobile devices offload computational overhead to network edges or nearby devices. Notwithstanding the rosy outlook, a number of issues have to be addressed since there is a fundamental trade-off between computation and communication. In fact, when one tries to offload computational overhead, one should investigate the overhead of the communication link. Fig. 4 depicts computation offloading scenarios where a VR headset assigns linear computations to four edges with various offloading schemes [21]. Since the reduction in the communication load over the network increases the computation overhead at the edges, we need to come up with a proper offloading scheme optimized for the computation capability and network conditions. One thing to note is that the wireless channels experienced by devices and their computing capabilities are quite different among devices so that there is a large variation in the task completion time. Particularly, this problem is critical for the ultra-reliable and low-latency communications (URLLC) services since the task completion time is determined by the device with the worst channel condition or straggling/malfunctioning devices (e.g., edge failure in Fig. 4). For the seamless integration of communication and computing, the following approaches can be considered.
Communication-aware pervasive computation design: To alleviate the difficulties caused by the channel uncertainty, we need to exploit the mmWave and THz frequency bands. If sufficient bandwidth is provided and LOS channel conditions are guaranteed for every computing node, wireless links will not be a serious bottleneck. Obviously, this ideal condition cannot be guaranteed always, so we need to come up with an environment-aware and communication-friendly computing mechanism. For example, the communication load can be reduced by quantization, compression, and sparsification of the transmit information as long as it does not severely degrade the task quality. For special cases where the (weighted) average of the computation tasks is needed, we can design the system such that the computation results are added during the wireless transmission of the same time/frequency resources. This socalled over-the-air computation can significantly reduce the communication and computation overhead.
Collaborative and robust coded computation offloading: A naïve approach to exploit multiple computing nodes is to split the whole computation task into multiple pieces and then distribute them equally to multiple nodes. This split-based offloading will not be effective in the presence of link failure and straggling/malfunctioning devices, since the whole task can be completed only when all subtasks are finished. One way to avoid this problem is to replicate a part of computation tasks and assign them to the multiple devices.
One option to overcome difficulties caused by the channel heterogeneity, link failure, and straggling/malfunctioning is to apply coding techniques to distributed task allocation. By introducing redundancy in distributed tasks, coded distributed computing can be more resilient to the link and edge failures. For example, if computation tasks are designed using maximum distance separable (MDS) codes, one can recover the straggling tasks from other completed tasks [22]. In Fig. 5, we show the task completion times for offloading linear computation tasks over two different channel scenarios. We adopt the computation time and communication time models in [23] with a task (matrix) size [TeX:] $$l=1 \times 10^7$$, computation intensities [TeX:] $$\eta=[300,500]$$ (CPU cycles/bits), CPU frequency [TeX:] $$f=1 \times 10^9$$ (Hz), transmission bandwidth W = 2 (MHz), channel noise power density [TeX:] $$N_0=1 \times 10^{-19}$$ (W/Hz), transmit power P = 0.1 (W), and an exponential computation delay with a straggling parameter [TeX:] $$\mu=\left[1.8 \times 10^7, 1.5 \times 10^7\right]$$, for two different scenarios, respectively. As illustrated, one can achieve the better computation-communication trade-off by assigning coded computation sub-tasks to computing nodes.
Task completion times when a user offloads linear computation tasks to four edges over two different channel conditions (Scenario 1: LOS channel at a distance of 15 m; Scenario 2: NLOS channel at a distance of 300 m).
Scalable and over-the-air deep learning: To perform a largescale deep learning (DL) task over wireless networks, we must decide how to allocate the tasks among the cloud, network edges, and the mobile devices [24]. For instance, network edges perform DNN training, and the mobile device performs the inference using the trained model from edges. To perform the flexible DL using abundant network resources, we can consider the following: 1) Federated learning, which trains local models on each device and then passes updated parameters to the central cloud for a global model; 2) distributed learning, in which each device communicates directly with its neighbors; and 3) split learning, in which the device or the central cloud trains each split model individually.
These DL techniques can be combined with over-the-air techniques [25] to effectively aggregate large-scale distributed data in ultra-low-latency scenarios. In addition, a scalable approach can be applied to resource-constrained mobile devices to reduce communication costs by leveraging gradient clustering, correction, and quantization methods.