Good luck building a virtual, open RAN – there's no such thing

Non-Spanish or Catalan speakers who travel to Barcelona for Mobile World Congress (MWC) choose from a bewildering menu when they sample the obligatory late-night tapas. Telcos must experience similar confusion in selecting options and suppliers for a radio access network (RAN). Nearly all operators used to get the same traditional meat and potatoes, served up by an Ericsson, Huawei or Nokia. These days, the list of ingredients and flavor combinations is phone directory-long. And none tastes how it is described.

Operators seeking alternatives to the familiar trio of RAN giants face disappointment at worst and compromise at best. It all starts with the silicon that hosts the RAN software, and specifically the baseband functions classed as "Layer 1" by industry lingo. In the traditional meat-and-potatoes variety, this would be tailored to the needs of that software, developed in-house by a RAN giant and any silicon partners. In a new-fangled "virtual" RAN (or vRAN), the hardware and software would be divorced, allowing the software to hook up with any general-purpose processor on the market. But this nearly always means Intel.

The chips behemoth of the Western hemisphere has established an effective monopoly in the nascent vRAN market. If it is allowed to survive, and vRAN takes off, operators could be trading a RAN oligopoly for dependency on Intel, famously blamed for a 5G crisis at Nokia after failing to deliver chips on schedule several years ago. "If operators want openness and diversity and lots of different suppliers, why would it be good to have a monopoly in Layer 1 compute?" said Tommi Uitto, the head of Nokia's mobile networks business group, during an interview with Light Reading at MWC.

Figure 1: Nokia's Tommi Uitto addresses analysts and reporters at MWC.
(Source: Nokia)

That monopoly already looks hard to upset. Branded FlexRAN, the Layer 1 reference design that Intel usually provides with its chips is bound tightly to x86, the umbrella term for an entire family of general-purpose processors. Intel effectively owns the x86 platform, and its only significant licensee is AMD, which spells out the risks whenever it files an annual report with the US Securities and Exchange Commission: "Intel has been able to control x86 microprocessor and computer system standards and benchmarks and to dictate the type of products the microprocessor market requires of us."

At least FlexRAN can easily be used with AMD processors, right? Not necessarily, according to Joel Brand, the senior director of product marketing for Marvell, another chipmaker. "FlexRAN not only runs on x86 but also requires the AVX-512 instruction set, and it is only supported by Intel," he told Light Reading last October. "You can't run FlexRAN on an x86 by AMD."

This may be changing. Reports last year suggested that some of AMD's latest processors do include support for AVX-512, and Sachin Katti, the head of Intel's network and edge group, insists FlexRAN is deployable on AMD chips. While his critics complain about Intel's vRAN monopoly and ability to "lock in" its customers, Katti says the real problem lies with rival silicon from the likes of Marvell, Nvidia and Qualcomm. None of it can support a genuinely virtual and open RAN, he argues.

Arm wrestling

The criticism stems from all three companies' adoption of a technique called inline acceleration. It is intended to address the performance and power-efficiency problems that arise when a general-purpose processor is used to support very demanding RAN software. It does that by moving Layer 1 functions off that processor and onto another chip hosted on a separate accelerator card. That's an expensive workaround, according to Katti, and it also means backing away from virtualization and openness.

"If you are going to have a big chunk of your software hard-coded into your hardware then you are not cloud-native," he told Light Reading at MWC. What does this mean from a practical perspective? For one thing, operators would not be able to upgrade software with the same Kubernetes toolset used everywhere else (Kubernetes being the de facto standard for managing cloud-native software). People would have to be trained to deal with that, said Katti, and operators would be tied to one supplier. "You are locked into a vendor because that is a custom accelerator."

What unites Marvell, Nvidia, Qualcomm and others is their use of blueprints developed by Arm, a UK-based company (currently owned by Japan's SoftBank) known mainly because its designs underpin most smartphone processors and are admired for their power-efficiency. This whole debate about silicon options and lock-in, then, effectively pits the Intel camp against the Arm camp, with Intel pushing for heavier reliance on its central processing units (CPUs) and Arm's licensees trying to capture vRAN ground. But hardly any telcos have come out strongly in favor of one system over the other.

Figure 2: Vodafone's Yago Tenorio shows off some of his company's tech in Barcelona.
(Source: Iain Morris/Light Reading)

One executive concerned that neither side offers a truly open vRAN is Yago Tenorio, the network architecture director of Vodafone. Previously critical of FlexRAN, Tenorio was just as scathing at this year's MWC about companies promoting Arm-based inline accelerators. "It is the same in both camps," he said during an interview. "With inline, you are depending on a piece of silicon that is taking care of Layer 1 and you also need to adapt your software to the instruction set that comes with it. In both cases, I think there is a certain amount of lock-in."

As things stand, Tenorio is also unconvinced that inline would offer better performance. Intel claims to have tackled the Layer 1 drawbacks of general-purpose processors through alternative methods of acceleration that keep the CPU more heavily in the mix. Lookaside, Intel's initial method, also introduced separate silicon, using this to process some of the Layer 1 functions. A newer "integrated" approach puts the accelerator and the CPU on the same chipset, says Katti, and has cost and performance advantages over inline.

On that, Nokia disagrees. While still an Intel customer, Uitto belittles integrated as a renaming of lookaside and says all his trials have shown inline delivers better results. Nokia, accordingly, has picked its team and is basing its entire vRAN strategy around the purchase of inline accelerators. One of its main suppliers appears to be Marvell, from which it buys identical silicon it can use to service traditional RAN contracts. Yet Tenorio is less ready to back a side. "The jury is out," he said, when asked if inline was superior.

Hitting the accelerator

The possibility of vendor lock-in through either system is something telco executives now want to avoid. The O-RAN Alliance, an operator-led group developing open interfaces, has recognized the issue within its Working Group 6, the bit dedicated to cloudification and orchestration. Its potential solution is something it calls the Acceleration Abstraction Layer (AAL). "Imagine if there is industry harmonization of instruction sets so that porting software from one piece of silicon to another is more straightforward," explained Tenorio.

A starting point would be to introduce a standard interface between Layer 1 and Layer 2, a separate part of the RAN software stack. Doing that would make the software less hardware-dependent and simplify the job for software developers, according to Panch Chandrasekaran, the head of Arm's 5G carrier infrastructure segment. "If you are an integrator and you have one Layer 1 implementation by Qualcomm and another by Marvell, then integrating those is a lot of effort," he said. "By standardizing that interface, you reduce the pain of integration for the different physical layer options you have."

Figure 3: Intel's share price ($) has suffered in recent years (Source: Google Finance)

Yet totally disaggregating Layer 1 hardware from software has proven difficult, according to Nokia's Uitto. And trade-offs between performance and supplier diversity may be impossible to eradicate. "When you use general-purpose silicon, you compromise and in exchange you are getting this very broad ecosystem of developers," said Marvell's Brand. "If operators are OK with a smaller number of providers, they will get optimized networks on optimized silicon. It's a choice."

Whatever success the O-RAN Alliance enjoys, its AAL initiative will do nothing to make x86 compatible with Arm, reckons Katti. "It doesn't solve the x86 to Arm portability problem," he said, noting there have been "decades of investment to optimize software for x86." It is a message that may jar with telco executives who are enthusiastic about AAL.

Providing feedback to Light Reading on condition of anonymity, one high-level telco source says the Intel rationale is probably that x86 – as a so-called complex instruction set computing (or CISC) architecture – includes many features Arm lacks with its RISC (reduced instruction set computing) alternative. But he also says Katti's remarks about non-portability imply Intel's technology is not as general purpose as it claims.

Figure 4: Global data center CPU market revenue share by vendor (Source: Counterpoint Research)

The nomenclature is arguably misleading if the platform is controlled by a single firm and has only one notable licensee. The foundational problem with virtualization is that the underlying world of general-purpose processors remains dominated by x86. A proxy is the market for data center servers, where these chips are commonplace. Figures provided by Counterpoint Research show that Intel's market share did fall from 80.7% in 2021 to 70.8% last year, but most of the gains were made by AMD, whose share grew from 11.7% to 19.8% over this period.

A truly vibrant and open vRAN sector would sport options besides Intel and AMD for Layer 2 and Layer 3, and these other parts of the RAN software stack currently make Layer 1 look super-competitive. "The fundamental question you have to ask is whether it is worth investing to port software when you have spent years making it work on x86," said Katti. But unless those investments happen, the telco mission of cultivating openness and diversity could ultimately be in vain.

Related posts:

— Iain Morris, International Editor, Light Reading