The goals for the so-called 5G Radio Access Network (RAN) are lofty indeed and have been discussed at length by industry experts. What has received far less airtime is the question: “What exactly is the best path to 5G?” This article lays out some of the challenges of the 5G RAN and ways in which ideas can be implemented in hardware — both for prototyping, which needs to happen over the next three years, and ultimately for production deployment, which is slated to commence in 2020.
5G: Evolution, revolution, or both?
The goal of 5G is to provide a 1,000x increase in capacity, supporting 100+ billion connections with data rates up to 10Gbps and less than 1ms latency. However, these new networks will not just support the fastest links and fattest data pipes; they also aim to improve upon the capabilities of current networks. For example, today’s wireless networks lack support for the low data rates and long battery life required for M2M (machine-to-machine) and sensor-type technologies.
Developing 5G networks that meet these goals will require a combination of existing systems such as LTE-Advanced and WiFi, combined with revolutionary technologies designed to support new uses such as the Internet of Things (IoT), augmented reality, immersive gaming, and UHD (ultra-high-definition) streaming video.
Major innovation is also needed at the lowest levels to accommodate broad requirements for both video and augmented reality. The needs of M2M networks will drive innovation in the physical layer, air interface definition, and control plane structures.
New frequency bands
5G will see some of the spectrum below 6GHz being re-purposed for use with newer technologies, particularly for non-line-of-sight (NLOS) requirements. Existing cellular bands will be augmented with new spectrum allocations above 6GHz that are able to supply much wider contiguous spectrum. Additionally, carrier aggregation techniques will be used to combine chunks of spectrum that are not co-located within the same band to further improve peak data rates. The core bands will provide up to 100MHz of instantaneous bandwidth, and the new extended bands will provide contiguous chunks of spectrum with as much as 500MHz in bandwidth — perhaps more.
Figure 1. Existing and likely 5G frequency bands.
Release 12 of the 3GPP standard, slated for freeze in 2015, provides for early massive MIMO (multiple-input and multiple-output) systems. These systems take active antennas to a new level. Large arrays of radiating elements (16×16 to 256×256 MIMO) require horizontal and vertical beam-forming to significantly increase capacity and coverage. In turn, massive MIMO requires significantly more processing power.
Advanced physical layer
Current 4G OFDM (orthogonal frequency-division multiplexing) air interfaces deliver high-speed data with limited support for low-power M2M communications. As a result, air interface technology and the 5G physical layer will be augmented using new bands of spectrum as they become available. Many new candidate air interfaces are being considered to provide support for sub-1ms latency with 10Gbps throughput. Other interfaces that can cater to the needs of simple sensor data transmission will not require such low latency or high data throughput, so it’s likely that 5G will not employ a single air interface technology. Equipment will need to support multiple air interfaces — potentially simultaneously.
In addition, the physical layer will require new coding and modulation schemes, protocols, and framing structures brought about by disparate end-user requirements. The 5G infrastructure must automatically determine the type of channel needed, and adapt based on conditions (such as precipitation) or moving objects (such as trains, airplanes, or cars) affecting line-of-sight (LOS). Cognitive radio techniques and advanced adaptive coding and modulation schemes will allow equipment to provide the best possible connections.
Existing basestation architectures consist of a cabinet housing radio units, power amplifiers, and baseband cards along with control and backhaul access. More recent architectures move the radio units to the mast, adjacent to the antennas, to eliminate lossy coaxial feeder cables and improve energy-related OpEx (operational expenditure).
New architectures including Cloud RAN and Virtual RAN take a more centralized approach for greater CapEx (capital expenditure) and OpEx savings. Centralizing baseband processing and backhaul functions to serve many hundreds or thousands of remote radios enables the use of GPU-centric server farms with localized data-center processing at the edge. This change places significant challenges on the fronthauling aspects of the networks, where the data from many hundreds of radios must be transmitted to data centers over various media (e.g., copper, fiber, over the air). 5G infrastructure will also push in other directions — including core virtualization components such as Software Defined Networks (SDN) and Network Functions Virtualization (NFV) — resulting in a more software-centric, server-based architecture that allows use of commodity servers and distributed processing.
Implementing Massive MIMO today
The benefits of Massive MIMO are undisputed, but the cost of implementing Massive MIMO is enormous due to the computational burden involving large matrices and linear algebra for beam-forming calculations for each antenna. As a result, Massive MIMO will hugely increase both connectivity and signal-processing requirements. High-speed connectivity is required between the digital front-end (DFE) processing and the analogue domain — many of the data converters are migrating to JESD204B — and between the baseband processing and the radio processing, which require some form of serial transceiver. DSP for DFE and beamforming algorithms demands wide bandwidths and high sample rates, necessitating agile, high-performance signal-processing.
Figure 2. Massive MIMO concept leveraging FPGAs and APSoC.
Today, Massive MIMO antenna algorithms can be realized with current technology, as shown in Figure 2, but as Massive MIMO systems scale to larger and larger arrays of antenna elements, greater levels of integration will be required. This will be made possible by future device generations.
FPGAs and SoCs enable massive connectivity and capacity
The capacity and latency goals that 5G demands will have a knock-on effect to the requirements of the infrastructure equipment. 5G systems must support massive connectivity and massive capacity that can only be served through the use of high-throughput communications. including 10Ge, 40Ge, PCIe, and future evolutions of CPRI. Capacity increases will come from new modem architectures, advanced radio technology, and new modulation schemes — all of which require huge increases in signal processing capabilities.
FPGAs have long been used in wireless infrastructure equipment due to their high performance, which permits rapid implementation of complex signal-processing algorithms. For example, the latest Xilinx 20nm UltraScale FPGA devices can support over 8 TMACS and more than 5Tbps of serial transceiver bandwidth. Xilinx All Programmable SoC devices couple high-performance FPGA fabric with a fully-integrated processing subsystem based on dual ARM Cortex A9 MPCore processor’s, which can be used to efficiently implement higher layers of the complex 5G protocol stacks.
Fronthauling is an evolving market that is driven heavily by the centralization of baseband processing, which — in turn — drives the need for IQ data fronthauling by wireless, copper, or fiber media. Current connectivity standards exist in the form of CPRI and OBSAI. Figure 3 shows a state-of-the-art CPRI aggregator implemented in an FPGA.
Figure 3. FPGA implementation of CPRI fronthaul aggregator
(Click here to see a larger image.)
5G is likely to have a different implementation for some processing elements. For example, Layer 1 baseband processing may move to the radio to reduce overall payload bandwidth, which drives greater integration in the radio domain. Whether Layer 1 is integrated within the radio or not, 5G development will continue to focus on baseband processing and the radio and on associated fronthauling technologies.
Advanced physical layer evaluation tool flows
Development of the physical layer for 5G is underway with many candidate technologies. Evaluating the relative merits of new candidate air-interface technologies and their associated Layer 1 processing needs is best done with FPGAs, which enable rapid implementation of required algorithms and interfaces. The inherent re-programmability of FPGAs permits rapid design changes to demonstrate improvements or to add features with very little schedule impact.
High-level synthesis (HLS) tools ease development of advanced 5G algorithms. For example, Xilinx’s Vivado HLS enables algorithm developers and system architects to design in C/C++ and then synthesize to RTL as shown in Figure 4. Popular third-party tools — including MATLAB and Simulink — can also be used for front-end design.
Figure 4. 5G IP creation using C/C++, RTL, or System Generator for Simulink
(Click here to see a larger image.)
With the advent of All Programamble SoCs, such as the Xilinx Zynq SoC family, ARM processors are readily available for implementing scheduling and other higher-level protocols. All of the normal tools in the ARM ecosystem are also available for these designs.
5G prototyping platforms
5G standards do not yet exist, but companies are still keen to start prototyping. Creating custom hardware and software to implement 5G functions is both time-consuming and costly. BEEcube has developed a range of 5G prototyping hardware that takes much of the pain out of developing such systems. This means that system architects can focus their efforts on where they can add value — with the development of new 5G candidate technologies.
BEEcube provides several platforms (see Figure 5) that can easily be tiled together to create huge amounts of DSP processing power (>100 TMACs) and large amounts of optical connectivity (>10Tb/s) for CPRI aggregator fronthauling designs. Each platform supports VITA-57 FMC analog cards to be fitted easily for direct RF sampling or for interfacing 1GHz of bandwidth to a 60GHz transceiver. BEEcube also delivers of all the required 5G interfaces, enabling system designers to focus on developing the algorithms without getting stuck on the interface standards.
Figure 5. BEEcube 5G development platforms.
5G production technology
FPGAs will be used for prototyping 5G wireless infrastructure over the next few years. There really are no other alternatives. However, when it comes to deployment technology, cases can be made for both ASICs — the traditional choice for high-volume manufacturing — and FPGAs.
The decision of whether to keep a design in an FPGA or migrate to an ASIC for production is a question of economics. With more serial transceivers, DSP slices, block RAMs, DLLs, PLLs, processor sub-systems, memory interfaces, PCIe interfaces, and other blocks, the FPGA’s hardware penalty for re-programmability continues to diminish. In parallel, the risk of severe ASIC design bugs increases exponentially as overall 5G system complexity increases.
Designers once agonized over whether to load software into ROM or PROM. Today, no-one stores program code in a non-re-programmable memory. Now it all goes into Flash memory, which can be upgraded and reprogrammed in the field. The same trend is happening with hardware. It is likely the 5G wireless infrastructure OEMs will bet on programmability to reduce design risk and speed time to market.
The race is on to solve many 5G technical challenges. We’re still five years away from commercial deployment, but — as standards begin to firm up — many companies need to prototype these emerging algorithms and applications now. Xilinx FPGAs and Zynq SoC devices — coupled with commercially available 5G prototyping platforms such as those from BEEcube — can save significant development time versus the development of custom prototyping platforms. These tools allow system architects and designers to get on with the job of finding the best architectures and algorithms, rather than spend their time architecting the platform on which to prototype.
As we look to 2020 for widespread 5G deployment, it is likely that most OEMs will sell production equipment based on FPGAs and All Programmable SoCs. The hardware complexity of 5G’s physical layer is just too challenging to guarantee that ASIC implementations will be free of severe hardware bugs. Keeping the hardware soft will be the wise path chosen by the smartest OEMs.