Intel’s domination of the general-purpose processors (GPPs) found in open and virtual radio access network (RAN) technology has long been awkward for champions of the concept. Around this time last year, the chipmaker claimed a 99% share of the virtual or cloud RAN market – great for Intel, but not ideal for those arguing this market will be more open and competitive than the traditional RAN sector controlled largely by Ericsson, Huawei and Nokia.
Diehards reject the comparison. The open and virtual RAN market remains tiny, for a start, and constitutes an alternative per se. Anyone can theoretically develop RAN software for an Intel GPP, using common programming languages and technology platforms. It’s not like that for an application-specific integrated circuit (ASIC) built by Ericsson, fusing hardware and proprietary software from the outset.
Yet neither point is very sound. Cloud RAN is forecast to be much bigger in future. If proponents have their way, it will eventually be the preferred approach. In the absence of GPP alternatives, operators would then look heavily dependent on Intel. And even if Intel’s GPPs are more open and programmable than ASICs, its x86 architecture is not an industry standard but largely the preserve of Intel. AMD is the only other big chipmaker to use it.
All of which makes developments last year seem rather exciting. Eager to foment GPP competition to Intel and x86, numerous RAN stakeholders have swung behind Arm, a UK-based chip designer known mainly because its blueprints underpin most of the world’s smartphones. Arm, unlike Intel, does not make chips but instead licenses its technology to companies such as Qualcomm and Nvidia. Several of these licensees are now using Arm’s blueprints to build GPPs for use in the RAN.
Arm’s RAN army
A notable update came at the end of 2023, courtesy of Nokia. For a cloud RAN, the Finnish vendor would conventionally divide its software between Marvell Technology – whose custom chips are used for Layer 1, the most demanding category of functions – and Intel’s GPPs, for the less challenging Layers 2 and 3. But during lab trials last year, Intel was replaced by a relatively obscure player called Ampere Computing.
Heavily financed by Oracle, and working in partnership with server maker HPE, Ampere has developed an Arm-based GPP suitable for rollout in the RAN. Substituting it for Intel would not seem to require any major rework by Nokia, which says it can effectively deploy the same code for Layers 2 and 3 on either x86 or Arm-based GPPs.
Interest in Ampere also comes from Vodafone. Last October, at the FYUZ industry conference in Madrid, the UK-based telco said it was involved in tests of Fujitsu’s RAN software on Ampere’s GPPs. “We are testing with Fujitsu software because they were more open and more available than the others, but more vendors will come,” said Santiago Tenorio, Vodafone’s network architecture director, at the time. “I know Ericsson and Nokia and Samsung are all paying attention to this because they really want competition to Intel.”
Ampere, though, is not the only Arm product in contention. The Grace part of Nvidia’s Grace Hopper “superchip” is also an Arm-based GPP. Far more prominent outside the confines of the RAN is Graviton, a chip built with Arm blueprints by Amazon and already deployed in the data centers of its AWS cloud subsidiary. When Counterpoint Research calculated the market shares in 2022 for suppliers of central processing units (CPUs) to data centers, AWS, with its 3.16%, was the third-biggest player after Intel and AMD.
As it did with Ampere, Nokia has already substituted Graviton for Intel during cloud RAN trials highlighted last year. An update on this activity is expected at the forthcoming Mobile World Congress in Barcelona. In a bid to cultivate an Arm ecosystem, AWS has also been helping Arm to port and adapt software written for x86. The RAN stack is one target, said Ishwar Parulkar, the chief telecom technologist for AWS, during an interview with Light Reading last year.
Capacity boost
Telcos are drawn to Arm partly because it is lauded for its energy efficiency. AWS has made an especially big deal about this, insisting Graviton’s price-performance ratio is 40% lower than x86’s for most workloads. The downside is that Arm-based GPPs tend to be less powerful. To compensate, Arm-based RAN equipment might have to include more “cores,” the building blocks of a CPU. And that would risk driving up costs.
Developments in vector processing could help. This would essentially allow the CPU to handle an array of data in one go and therefore differs from scalar processing, which deals with one element at a time. That makes older CPUs based on scalar processing less suitable for the computationally hungry part of the RAN. An instruction set called AVX-512 ticks the vector-processing box for x86. For a long time, however, there was not a commercially available Arm equivalent.
This is now changing thanks to the inclusion of SVE2, an Arm instruction set, in newer Arm-based GPPs. “That is basically vector processing and is very suitable for Layer 1 processing and, when they get that in, we think we can squeeze some good capacity out of an Arm system,” said Matteo Fiorani, the head of Ericsson’s distributed unit and infrastructure business, in September last year.
Unlike Nokia, the Swedish vendor says it wants to keep as many of the Layer 1 functions as possible on the GPP, rather than putting them on a custom chip used as a so-called “accelerator.” This strategy would seem to make SVE2 more important to Ericsson than it is to Nokia. Nevertheless, writing hardware-agnostic code, deployable on any chip, is much harder to do for Layer 1, according to numerous experts. “Tweaks and optimizations” would be necessary, said Fiorani.
Hardware agnostics
But in December, at UK offices in Bristol, a smaller RAN software developer called Parallel Wireless showed off technology compatible with both x86 and Arm-based GPPs for Layers 1 to 3. It has managed this partly by observing and learning from the hyperscalers that now use both x86 and Arm-based chips across their various workloads, said Nicolas Scheidecker, the UK head of research and development for Parallel Wireless.
Scheidecker concedes that accelerators might still be needed in the most advanced 5G networks based on “massive MIMO,” which effectively crams dozens of antennas into basestation equipment. But he also thinks AVX-512 and SVE2 could address many of the Layer 1 performance issues associated with GPPs. And Parallel Wireless has now tested software with chips from a range of companies, including Ampere, Marvell, Nvidia and NXP.
This all sets 2024 up to be a potentially game-changing year for cloud RAN. Yet Intel is likely to remain by far the biggest player for the foreseeable future. Arm’s ecosystem is immature by comparison. And Intel may have a point when it argues that no two Arm-based chips are alike. “Every variant of Arm is different and so Ampere’s Arm product is different from AWS’s Arm product,” said Sachin Katti, the general manager of Intel’s network and edge group, in September. “And frankly, I don’t think software you write for Ampere is easily portable to Arm running on AWS.”
Regardless, the chief technology officer of one Tier 1 telco was recently heard bemoaning the open and virtual RAN concept on the grounds that it would make him reliant on a single chipmaker. Vodafone was evidently motivated enough to involve itself in trials of an Intel alternative. The foundations of an Arm-based system are finally in place. This year will show if the industry can quickly build on them.
