Vodafone UK pushes open RAN past faulty towers, with NEC missing

Torbay in southwest England is famous for beaches, detective novelist Agatha Christie (born there in 1890) and Fawlty Towers, a 1970s sitcom about a cantankerous hotelier. Appropriately enough, Vodafone UK chose it for its initial experimentation with open RAN, a newish technology concept that has provoked concern about faulty towers of a different kind. Using general-purpose equipment and combining vendors, as open RAN proposes, has brought risk of performance problems at mobile masts. Yet Vodafone appears sufficiently happy with its Torbay trials to have shifted up a gear.

The plan, first announced two years ago, is to extend open RAN to thousands of sites in Britain's west currently served by Huawei. Under government rules, mobile operators must remove any 5G products bought from the controversial Chinese vendor by the end of 2027. Vodafone's rollout of a "golden cluster" comprising just 16 trial open RAN sites in Torbay was announced more than a year ago. It has now given itself less than four-and-a-half years to install no fewer than 2,500.

In theory, this should be relatively easy. UK telcos first launched 5G in late 2019 and today boast widespread coverage. BT, the incumbent, claimed 5G population coverage of 68% in its last annual report (for the year ending in March 2023) across a network of about 19,000 mobile sites. But it's clear Vodafone's open RAN project has not advanced quickly, and many of the technical concerns have not been fully addressed.

What's more, Huawei's technology goes well beyond these 2,500 sites. Back in early 2019, when UK authorities were still pondering how they should deal with Chinese vendors, Vodafone revealed at a press conference that Huawei supplied RAN technology for about 32% of its footprint, or roughly 6,000 sites at the time.

Executives have previously acknowledged their interest in extending open RAN to other parts of the Huawei footprint not covered by today's plan. And in its latest statement, Vodafone promises to have open RAN installed across "at least" 2,500 sites by the end of 2027. But the gap between 2,500 and 6,000 is huge.

Huawei might not supply 5G products to all these sites, but operators typically use the same vendor for multiple generations of mobile technology. While complying with government rules, Vodafone will probably want to avoid retaining Huawei for older standards alongside a different 5G provider. There is a growing possibility it will turn to Ericsson, its main 5G vendor, to replace a chunk of the Huawei estate.

Massive MIMO mood swing

Perhaps topping the list of technical concerns is massive MIMO. This antenna-rich 5G technology is now widely used in commercial deployments, and yet numerous stakeholders have flagged up the inadequacy of the 7.2x category B specification aimed at massive MIMO by the O-RAN Alliance, a telco-led group behind open RAN. Following complaints by Ericsson, Orange, Qualcomm and others, the O-RAN Alliance signed off on two optional "operation mode" improvements at a June meeting in Osaka. Neither looks ideal.

In short, the problem with 7.2x was the decision to put critical uplink functions into the distributed unit (DU), a server box. While this simplified the design of any radio unit (RU), it threatened an explosion in traffic between the DU and the RU, one that could potentially hurt performance in areas where fiber is not ubiquitous.

The class (or operation mode) A fix, sponsored chiefly by Ericsson, is to move the interference-addressing equalizer and other uplink functions from the DU into the RU. But Orange and Qualcomm want the equalizer kept in the DU. Their class B alternative aims to address performance problems by instead carrying out channel estimation, another important function, in both the DU and the RU. In plain vanilla 7.2x, it is limited to the DU.

The Osaka compromise is messy. Radio vendors can manufacture products based on either class A or class B and still be considered O-RAN-compliant, but DUs must include equalizers, even if a class A deployment makes them redundant. That effectively means class A is costlier than it needs to be. As for class B, doing channel estimation twice could lead to interoperability problems if the RU vendor is different from the DU vendor and their algorithms do not match, according to Kim Larsen, a former chief technology officer within the Deutsche Telekom group.

Vodafone declined to say which approach it would follow, describing this as "commercially sensitive" in an email. But massive MIMO is clearly mentioned in a press release about the commercial rollout issued by Samsung, the South Korean company supplying both DUs and RUs to Vodafone. Intriguingly, it backed class B during the O-RAN Alliance dispute, while Vodafone preferred class A, according to a reliable source. Samsung's current range of products, though, is undoubtedly based on the original 7.2x spec.

Interoperability concerns could explain why no other DU or RU vendors seem to figure in the current plans. The original premise of open RAN was to ensure that one supplier's RU could pair with another's DU, and Vodafone's initial announcement in June 2021 identified Japan's NEC as a massive MIMO supplier. The latest update names every other vendor that was originally called out, with NEC conspicuous by its absence. Unless it is used, or another vendor of the most important components is introduced, the network Vodafone builds will not look very open.

Accelerator wars

The other big technical concern is about the use of general-purpose processors (GPPs) in DUs. Traditional networks rely on purpose-built silicon for RAN processing. GPPs would support the virtualization or cloudification of the network, but they are widely seen to be less energy efficient. The industry's workaround is to offload the most computationally demanding RAN functions – categorized under the Layer 1 label – onto other silicon chips, dubbed accelerators. But there are trade-offs to consider.

Vodafone has confirmed that it will initially use an Intel-backed acceleration technique called lookaside. With this, only a couple of Layer 1 functions are handled by the accelerator, leaving the GPP heavily in play. This should make it easier to manage all the network functions with a single software platform. But critics fear performance will suffer because of a GPP's unsuitability for RAN processing.

The alternative, called inline, moves the entire Layer 1 stack onto an accelerator built with customized silicon. The case for inline is largely based on claims about superior efficiency and performance. Because it is normally delivered on a card separate from the main server, an operator could also add capacity just by slotting in new cards. By contrast, Intel's move to put accelerators on the same motherboard as the GPP would force Vodafone to buy a whole new server.

Inline remains optional for Vodafone because Samsung is hedging its bets. Besides writing software that works on Intel's GPPs and vRAN Boost accelerators, it has teamed up with Marvell, a chip developer behind inline. The trouble is that Samsung cannot simply lift the RAN software deployed on Intel's silicon and drop it on Marvell. The Layer 1 code would have to be completely rewritten to suit that accelerator. What's more, because GPPs are still being used for other network layers, Samsung would need multiple software toolkits – at least one for the GPPs and another for Marvell.

Regardless of which technique is used, the overarching problem is a lack of portability. Samsung could probably redeploy the RAN software matched with Intel's technology on AMD, which similarly backs lookaside and uses the same x86 architecture as Intel. But that's about as far as it goes. And switching between suppliers of customized silicon for an inline deployment looks much harder. Nokia, which buys Layer 1 silicon from Marvell, has confirmed that moving to another supplier would force it to rewrite code.

Santiago Tenorio, the network architecture director of Vodafone Group, was lashing out about this almost a year ago. "In simple terms, what the accelerator gives you is a new instruction set," he told Light Reading at the FYUZ event held in Madrid last October. "The instruction set, until we can standardize it, is proprietary for that piece of hardware. If you are Samsung, you can probably have one version for Marvell and one for Qualcomm and one for Intel, but it is not ideal and therefore the next wave of standardization needs to be on the Layer 1 interfaces, which are very difficult to abstract with middleware."

Yet there has been a total absence of visible progress on standardizing instruction sets in the last year. Various industry executives and experts that Light Reading has spoken to over that period seem to regard the tight coupling of Layer 1 hardware and software as a fact of life, as immutable as the laws of physics. If that's so, then openness has its limits.

Related posts:

— Iain Morris, International Editor, Light Reading