Virtualization is almost as old as commercial computation, dating back over half a century to IBM’s development of virtual machines running on its monolithic mainframe hardware. The motivation then was to extract maximum use of very expensive and scarce computers by allowing each one to run multiple operating systems and thereby emulate other machines with their own workloads via time sharing.
Now virtualization has, in one sense, been turned upside down, because while 50 years ago it was motivated by the high cost of hardware, it is now driven by the fact general purpose CPUs have become so cheap. This has made it increasingly attractive to migrate network functions away from dedicated chips (ASICs and FPGAs) onto COTS hardware based on processors such as Intel’s x86 range, despite sacrifices in performance, energy efficiency and increased complexity.
Other benefits include scalability, since additional off-the-shelf components can be readily upgraded; freedom to cherrypick from multiple vendors because standard interfaces are adopted; and, especially for 5G, ability to unify all parts of the infrastructure including core, backhaul, fronthaul and RAN. Many of the virtualization benefits and challenges associated with 5G are shared by other technologies, but some are unique.
In the case of the backhaul there is more overlap with the fixed broadband sector, especially regarding links over fiber, copper or cable. Broadband operators are mostly already embarked on virtualization in a process that will take around a decade to complete across the whole sector and that spills over into mobile backhaul.
Some of the findings by our sister research service, Rethink TV, in its report ‘Virtualization to capture 500 million fixed broadband customers by 2025’, are therefore applicable to backhaul.
This report forecast that the pace of broadband virtualization will increase almost everywhere across the cable and telco sectors, except in most of Africa. By 2025 there will be 504 million broadband subscribers globally connected via virtualized access infrastructure, 40% of the 1.26 billion worldwide total.
Underlying commercial drivers are overlapping with subtle differences. For telcos and cable operators, virtualization is being driven by intensifying competition as fixed broadband has become the major source of revenue. While that is also true for mobile operators, the major drivers there are the cost savings and infrastructure rationalization across the different reaches of the infrastructure.
The main differences not surprisingly lie in the RAN, where virtualization is really coming in only with 5G. That is the critical part or even the virtualization bottleneck, being essential to obtain the broader benefits that underpin the expanding diversity and use cases of 5G. In the RAN, virtualization has potential not just to cut costs and roll on integration with the backhaul and core, but also stimulate innovation.
First we should note that the path to RAN virtualization was laid by 3GPP in Release 15, with the split into three units, the Centralized Unit (CU), Distributed Unit (DU) and Radio Unit (RU). The CU in turn can be split into the user plane (CU-UP) and control plane (CU-CP), with the new 5G NR F1 interface linking the CU and DU. In principle then the DU, CU and RU can come from different vendors.
Just as importantly, this more distributed architecture also allows network functions to be separated from the underlying hardware, enabling NFV (Network Functions Virtualization). This in turn will help operators support multiple use cases via network slicing, while also being able to introduce COTS hardware.
Of course there are caveats and these have been played on by some of the traditional vendors, perhaps fearful of losing revenues to new providers of virtualized RAN or vRAN systems. Ericsson for example has been cautioning that vRAN will always have a larger form factor, be more expensive, consume more power and entail more complex system integration than dedicated RAN hardware.
All that is true today but some of those are costs that will be sunk, such as the systems integration, while other deficits such as additional cost and power consumption are likely to wither away over time, just as they have in other sectors. To be fair, Ericsson concedes the tide is flowing towards RAN virtualization and just wants to check the process.
But even then, Ericsson believes that commodity hardware in the RAN will still require a little more dedicated assistance.
On this count, Ericsson is reaching for a compromise by working with Nvidia, maker of graphical processing units and other acceleration hardware, to develop virtualized 5G radio access equipment. This will combine Nvidia’s GPUs with Ericsson’s own software, and also commodity CPUs. While Nvidia GPUs can be regarded as commodity components, they are dedicated to high performance applications involving large numbers of sequential operations that can be performed in parallel, including graphics rendering and machine learning.
The ultimate aim of the collaboration though is to establish an open platform that would allow components from other manufacturers to interoperate as well, so Ericsson’s end game is to establish a vRAN platform that will be ideally suited for its technologies to slot into and target new operators that already have some kit from other vendors. Over time, hardware assembled around this or other platforms will converge in cost, performance and size with dedicated components, which will then wither away.