Your browser is not supported. Please update it.

28 February 2020

Is general purpose silicon too slow and expensive for the vRAN?

Amid all the rising excitement about open RAN architectures, some large suppliers have taken an ‘if you can’t beat them, join them’ approach and are now aiming to turn open systems to their advantage by leading the charge.

Samsung and NEC have obvious motivations to leverage open architectures to improve on their very limited presence in access networks outside their home countries. Even Nokia is contributing code to the O-RAN Alliance and setting out a strong strategy to harness openness to lower its RAN costs, increase its market share, and refocus on higher margin software and services.

But the big two remain unconvinced. Ericsson has joined O-RAN Alliance but does not appear to be very active, while Huawei’s main contributions to open platforms have been for China Mobile’s internal projects. They are adopting the time-honoured response of incumbent mobile vendors in the face of challengers – remind operators just how much they have to lose if they implement their RAN poorly, and how much simpler, quicker and safer it would be to have a single vendor end-to-end.

Even suppliers which sign up for open platforms often adopt the latter argument in reality. Just because a network architecture allows for components and virtual network functions (VNFs) from different sources, there is no need for an MNO to take that route, argue some vendors. We have seen how companies like Cisco have disaggregated their architectures, only to tie everything to their own frameworks again.

The same will, no doubt, be attempted when open RAN deployments become big-ticket. But for now, operators which are keen to open up their networks and supply chains need to formulate good arguments against claims that open systems just won’t perform adequately at all, or that the new architectures will take too long to reach maturity for primary macro networks, to align with many MNOs’ 5G timelines.

Ryan Ding, president of Huawei’s carrier networks business group, was asked about open RAN in various conversations during what should have been the Chinese company’s pre-MWC showcase, in London last week. He insisted that Huawei was discussing all its customers’ and prospects’ requirements, including open systems, but steered well clear of the multivendor aspect of open RAN, claiming his firm could meet all an operator’s needs and was flexible enough to adapt to any particular need (and even price, was unsaid but implied).

Ding and his colleagues did say that a broader base of innovation and new entrants would be beneficial for all, and said Huawei was keen to work more closely with smaller companies for mutual R&D benefits. But he sounded a familiar warning from large mobile OEMs, that if open alliances’ specifications deviated from those of 3GPP, there was a risk of fragmentation and increased complexity. In that scenario, there would be lower economies of scale and greater complexity for operators to implement the new technologies. And for the foreseeable future, performance was likely to be compromised.

There may be plenty of FUD (fear uncertainty and doubt) behind such words, but there is also some truth. The computing industry’s journey from closed to open systems took a couple of decades, and many false starts and non-interoperable ‘open’ platforms, before Linux, and then the cloud technologies, became reasonably unified. The fear that open systems means fragmentation and lack of interoperability is a real one. Organizations like the Open Networking Foundation (ONF), Telecom Infra Project (TIP) and the O-RAN Alliance are making real progress in defining quasi-standards with open processes – greater alignment with one another, and with 3GPP, ETSI and IEEE, would turn these into a genuine 5G-class platform.

And the performance issues are real too. Virtualized, disaggregated architectures undoubtedly bring agility in terms of network topology and service flexibility. And software-centric networks deployed on COTS hardware will have, once they mature, lower cost of ownership than proprietary integrated base stations. But there is a huge remaining question mark. What silicon is required to power cloud infrastructure that has sufficient capability to support the demanding processes of a macro RAN? And what does that silicon cost?

Ericsson’s head of networks, Fredrik Jejdling, claimed at Ericsson’s ‘pre-MWC’ event that the dedicated custom ASIC [application-specific integrated circuit] processors that vendors design for their base stations cannot currently be matched by general purpose processors (GPP), when it comes to supporting advanced RAN functions. If a chip supplier surrounds the CPU with enough accelerators – FPGAs, GPUs (graphical processor units) or ASICs – to offload the most demanding tasks, the solution becomes very expensive and potentially power-hungry.

Nokia famously ran into this issue itself when it moved away from ASICs to an inhouse base station system-on-chip with heavy use of FPGAs (field programmable gate arrays), which proved expensive – the company is now developing a new generation of its Reefshark chipset, working with Intel.

In fact, Ericsson has been announced as a customer for Intel’s new baseband system-on-chip (SoC – see separate item), along with Nokia and ZTE; and it is engaged in a major R&D initiative with Nvidia, to test a virtualized RAN platform built around GPUs. However, it is clear that these approaches are not yet ready for the intensive processes of the distributed RAN.

Thomas Noren, Ericsson’s head of 5G commercialization, recently used dynamic spectrum sharing (DSS) as an example of an intensive, time-sensitive RAN workload which would require very high performance chips before it could be virtualized on cloud infrastructure. DSS allows an operator to run 4G and 5G flexibly on the same band and dynamically allocate spectrum between the two – it reschedules spectrum every millisecond, and it is very challenging for the hardware to handle all the time and frequency variables.

So far, such workloads require special-purpose, fully optimized processors, argued Noren. He said: “We have a multicore baseband architecture that allows us to do parallel processing; we have the capacity to introduce 5G and DSS on the same baseband. Our purpose-built baseband unit is the fastest, most power-efficient unit in the industry. That’s why we can do DSS.”

Jejdling made similar comments about Massive MIMO. “We spend significant amounts of R&D to deal with the compute power needed to handle massive MIMO and currently general processing cannot handle those types of workload,” he said. “So far we have not seen any technology that can deal with performance efficiency to the extent that ASICs can.”

In a disaggregated vRAN, most operators will have three elements – a central unit (CU), based in the cloud; a distributed unit (DU) closer to the cell site; and the radio unit (RU). Of these, the DU is the hardest to commoditize because it needs to support very low latency actions, and because it will be deployed in large numbers, its cost and power consumption must be very low. So while Ericsson, like other vendors, is virtualizing CU functions on general purpose hardware and Intel processors, the DU is seen as a more distant prospect.

Noren added: “If you look at all the available technologies, they are too expensive, too power hungry and too big to be effective compared to our purpose-built hardware for DU.”

It is a big concession, therefore, for Ericsson even to engage in an R&D project with Nvidia. Noren said: “Nvidia has a platform and development framework that we can potentially use. We will explore if we can develop a distributed baseband product (DU) with Nvidia GPU. We don’t have any committed product plans, but we think this is a very interesting idea.”

The view that ASICs remain cheaper and more capable for RAN was validated from a surprising source – Alex Choi, Deutsche Telekom’s SVP  of strategy and technology innovation and a prime mover in open networks, told LightReading: “The reason Ericsson, Huawei and Nokia are using more customized chipsets is that they give you a cheaper solution. If you have an ASIC for baseband processing, it is always cheaper than using a general purpose processor like an Intel processor.”

But like Ericsson with Nvidia, DT is always looking for new ways to implement networks, which may impact on cost and effectiveness in future. The German telco is currently testing a new technology developed by Intel, VMware and Cohere at its headquarters in Bonn and this has the potential to change the game. But it will need to be tested to destruction before operators are fully convinced, and getting it to the level required for a complex, highly distributed, virtualized RAN that will handle the MNO’s main traffic loads will be a long process.

“The performance, reliability and cost structure are not quite there yet,” Choi said. “The technology is progressing well but still not meeting all these key requirements from the operator side.”

Nobody expects these radical new approaches to be ready for prime-time overnight, and of course long cycles of development and testing will be required to build confidence. However, if this time stretches into several years, the risk for operators is that the moment will have passed – they will be ready, commercially, to deploy their next generation vRANs, but the open platforms will not be ready, and so they will have to fall back on the virtualized, but less open, solutions being readied by the big suppliers. These may well have open interfaces to the radio units. Just as large vendors are generally not excited by the revenues from small cells, and will work with smaller partners in that area as long as they keep control of the overall platform; so they could afford to let radio units from small vendors, based on open specs, into their RANs via open APIs, provided they still deliver the central and distributed basebands and the integration services.

However, Jejdling was unwilling to concede even this, telling LightReading: “What we see now is that with O-RAN specifications, the performance of those interfaces is not to the extent of the interfaces that are built on CPRI standards … It might be that certain use cases evolve. As and when the cost and performance benefits become a reality, we need to be there as a player.”

But even replacing semi-proprietary CPRI with O-RAN would not fulfil the dream of a fully open, interoperable network in which each of the vRAN elements could come from a different source; even each virtual network function (VNF) might be from a different vendor; and all this would be enabled by merchant silicon.

For that to happen within the timeframes leading operators have set for vRAN (within three years for the really ambitious ones), there will have to be support from the large OEMs as well as the start-ups. While Nokia and Samsung are showing willing, the big two are holding back. And they can significantly slow the pace for operators wishing to open up their RANs.

For instance, Ericsson has refused to support the particular options for 4G/5G coexistence that Vodafone was keen to deploy. Yago Tenorio, the operator’s head of network strategy, has been lobbying the industry to support three options defined by 3GPP – Options 4, 5 and 7. The 3GPP options specify how functions are split between central and distributed units, and how 4G and 5G base stations and cores coexist. Vodafone made its preferences clear when it issued an RFQ (request for quotation) last year for all its 150,000 European radio sites, insisting that vendors supported the O-RAN open fronhaul interface and its three favored splits.

But Ericsson will only support Options 1 and 2, returning to the anti-fragmentation argument. “We prefer to standardize the development across fewer options to harness and consolidate the R&D effort into certain standards,” said Jejdling. “We promote Option 1, and ultimately Option 2 in a [5G NR] Standalone scenario. And that is supported by the overall ecosystem, including chipsets.” Of course, chipset makers are unlikely to stick their necks out to support a certain option if they do not see large vendors being likely to buy the end products, setting up a vicious cycle in the macro layer – which contrasts with the small cell sector (see separate item).

Here, there are fewer incumbents, less interest from the big OEMs, and limited legacy installation, so there is far greater room for innovation and new approaches – which will start to bear commercial fruit in 5G because of its greater density and its emphasis on indoor and industrial use cases.

But for now, and for the foreseeable future, the networks remain macro-focused. This environment will rely on players with equal strength to that of Huawei and Ericsson to accelerate progress and push high performance, deployable solutions into the field. Intel’s announcements this week, and the support of Dell, VMware and others, provide hope for a full open RAN, especially if Nokia really puts its money where its mouth is and brings its radio expertise to bear on open specs.

The 3GPP splits in a disaggregated RAN:

The 3GPP standardizes eight options (13 including sub-options), with varying degrees of centralization, for the 4G/5G RAN. Only a few are likely to be supported on a large scale by operators and therefore vendors, but the role of industry alliances will be important to drive consensus and scale behind the options which will best support common 5G architectures.

The 3GPP is mainly focused on one high layer split, Option 2 (also called the F1 interface), for highly centralized RANs (with most functions virtualized in a central controller, often in the cloud). It is also working on two lower layer splits, with more functions pushed out to distributed controllers or the radio site. These two are

Option 6, with a MAC/PHY split, and Option 7, with an intra-PHY split, and offering three different variants 7-1, 7-2, 7-3.

The main 3GPP split supported by O-RAN is Split 7.2, while another industry alliance, Small Cell Forum, has aligned its own nFAPI (network function API) interface with Split 6. The two organizations are increasingly cooperating to ensure they work on platforms that can support both these options, to drive more scale into chips and equipment, and to support MNOs such as Rakuten Mobile, which want to deploy networks with multiple split options for different scenarios.

Some operators are taking the view that Option 6 is the strongest choice for affordable small cell deployments in lower capacity scenarios, especially indoors; while Option 7.2 will be superior for high capacity and ultra-reliable requirements, because it has strong support for multi-point transmission.

A common concern was that there are still too many options and that MNOs either do not know which to prioritize, or each will want a different combination – leading to fragmentation and lack of scale, with consequent risks for pricing and time to market. Some operators, like BT, are already driving the market towards a common hardware platform that could support all split options in software, to avoid the need to choose just one and to future-proof early deployments. In all cases, cost and availability of transport will be a key consideration.