5G is the first generation of mobile telephony that is heterogeneous by design, supporting both 3GPP and non-3GPP access technologies¹. 3GPP technologies include those that are carried by base stations built by traditional telephone equipment manufacturers, whereas non-3GPP technologies include Wi-Fi, a wireless broadband technology standard. However, coordination between the different technologies is only at a primitive stage.

Heterogeneous networks (HetNets), which use both 3GPP and non-3GPP access technologies, are already a feature of 4G, where a large fraction of the traffic is being carried by Wi-Fi, but they operate only in a post hoc and ad hoc fashion. 3GPP technologies offer features beyond those of Wi-Fi, they are though far more expensive to buy, deploy and maintain. In fact, in several countries, mobile services that are carried solely or predominantly by Wi-Fi, usually by using spare capacity on the routers of participating subscribers, already compete with traditional mobile telephony companies who rely on 3GPP-based equipment. In contrast to the step changes observed with 3GPP technologies, Wi-Fi has seen a continuous evolution in technology. This is due to the fact that the standardization of Wi-Fi is based on IEEE’s 802.11 standard² and is promoted by the non-profit Wi-Fi Alliance — and it is therefore outside the remit of the 3GPP remit.

Depending on the evolution of 5G infrastructure, Wi-Fi could become a major, or even the predominant, traffic carrier³. HetNet trials using both 5G and Wi-Fi have already been rolled out⁴ and are based on transport layer protocols, such as the multipath transmission control protocol. From a research and future-deployment perspective, the management of HetNets will thus be crucial. It may occur at the transport layer, by introducing, for example, network coding to blend different data streams with different throughputs, delays and reliabilities across networks, to provide delay gains⁵. Alternatively, the management of HetNets could be far closer to the fundamental physical layer. While mobile telephony and Wi-Fi have significantly different design principles, they have converged technologically in many aspects. In particular, both technologies operate orthogonal frequency-division multiplexing, which, in effect, already bonds multiple channels, each with different performance characteristics, by treating bandwidth as a collection of small distinct channels, or carriers.

Currently, one approach to provide reliability during data packet transmission is to use forward error correction, but at the expense of delay and throughput. While the signal-to-noise ratio of transmitted data is carefully controlled to allow for a finite bit error rate, high data rates ensure that many errors still occur. The need for robust, low-latency communications in 5G led 3GPP to explore the possibility of using polar codes⁶, which were recently introduced with a promise of low-delay operation. In particular, the current standard for 5G (ref. ¹) envisages the use of a refinement of polar codes, called check-aided polar codes⁷, for all control channel communications. The standard also anticipates the use of low-density parity-check codes, which were originally developed in the early 1960s⁸. In terms of electronics, the development of low-energy, low-delay decoders is a crucial part of 5G deployment⁹. Decoders have always been designed for use with corresponding encoding techniques. However, a recent development promises a universal decoder¹⁰ for the type of short codes and low bit-error rates that 5G appears to favour, which is compatible with all existing structured codes and even the random codes originally proposed by Claude Shannon¹¹.

Blending Wi-Fi and 3GPP technologies will change the architecture of 5G relative to its predecessors, as well as changing how the 3GPP radio base stations are managed¹². Far more of the wireless bandwidth is now being exploited and the number of base stations has increased. As a result, the legacy infrastructure of the local wireline network has grown relatively congested, negating the idea of unlimited wireline bandwidth.

In earlier architectures, only the last hop was wireless, and the wireline local network — generally termed the backhaul — was treated as a resource with unlimited bandwidth. In contrast, the use of wireless mesh topologies at the edge can reduce the dependency on wired infrastructure by allowing base stations to communicate among themselves wirelessly, without the need to be all tethered to an expensive wireline backhaul connection. An emerging approach in 5G is to implement self-backhaul base stations¹³, which are typically powered by high-carrier-frequency services, often termed millimetre-wave links¹⁴ (Fig. 1). Such high carrier frequencies rely on the aggressive use of a large number of directional antennas, often termed massive multiple-input multiple-output (MIMO)¹⁵.

Fifth-generation telephony (5G) is, by design, inclusive of different technologies, allowing traditional mobile services, wireline connections and Wi-Fi access, and communications between cars, Internet of Things (IoT) devices and even satellite. The 5G network uses a wide array of enabling technologies, such as hybrid access, millimetre waves (mm-waves), mesh protocols at the edge and among base stations, caching and MIMO antenna transceivers.

The use of wireless mesh connections instead of a wireline backhaul by the base stations suggests that this approach could potentially be scaled to a fully meshed wireless infrastructure¹⁶. A promising scheme towards this goal is virtual MIMO through cooperation, where users exchange information to self-assemble¹⁷. Recent results, however, illustrate the limitations of such an approach and the need for backhaul, when the cost of cooperation is taken into account¹⁸.

The terms ‘throughput’, which is defined in bits per second of data, and ‘bandwidth’, which is defined as the frequency spectrum assigned to a service, in hertz, are used almost synonymously in the literature and in common discussions. 5G is finally changing this by considering how bandwidth is being used: the available bandwidth at any one location for any particular user is more meaningful than the total assigned bandwidth. To address the related challenges, machine learning could provide the ability to detect and learn both the needs of users and the availability of resources in a fine-grained, dynamic fashion. For the first time in the evolution of mobile communications, the use of machine learning in network management has become an accepted innovation and is being considered in standardization by the International Telecommunication Union. This approach is quite distinct from the highly planned and hierarchical network management philosophy, which has its roots in the telephone system, from where the bulk of the mobile network operators and equipment manufacturers have originated.

Another change in infrastructure that 5G accelerates is the replacement of spectrum by storage. The fog — the term generally used to indicate the parts of the network that are closest to the end user — is increasingly being recognized as the largest part of 5G architecture. Importantly, the fog can provide storage in the form of caching (where content is strategically placed in a network to be readily available when demand for it arises¹⁹), availability of content, and remedial action when the content is not available.

5G is intentionally co-opting narrow markets with specific needs (verticals) that until now had separate, often bespoke solutions. The satellite industry and the automotive industry are prime examples of such industries. Some, such as the satellite industry, have rapidly and eagerly joined the discussion on 5G (ref. ²⁰). For others, such as the automotive industry, the role of 5G will heavily depend on its ability to provide the reliability and low delay that the industry requires. Much of the integration into different verticals (such as devices, often termed the Internet of Things) is based on network slicing²¹. Network slicing refers to using a part (slice) of the network resources to provide multiple different services to certain groups of users.

In the same way that throughput and bandwidth have often been equated, the inverse of delay has been commonly mapped to throughput. Throughput is the average number of bits delivered in a time unit, whereas delay is the time between transmission and reception of data in the correct order. For low-latency services, required for virtual reality, gaming and haptic applications²², as well as for navigation in automotive and other vehicular settings, network coding could provide resilience against errors in a time-limited time-frame rather than improve the long-run throughput. This may be incompatible with some of the approaches based on repetition of communication, which have underlined communications in mobile networks so far and which require high-throughput rates. Throughput and delay are actually two metrics that must be traded-off against one another²³.

Affiliations

1. Research Laboratory of Electronics (RLE), Massachusetts Institute of Technology (MIT), Cambridge, MA, USA
   • Muriel Médard

Corresponding author

Correspondence to Muriel Médard.