Nokia goes on warpath against Intel in cloud RAN

'Tis nearly the season of goodwill to all. But not much of it emanates from Oulu, where Nokia's little helpers design 5G networks, toward Santa Clara, the Christmassy-sounding location of Intel's headquarters. Several years ago, relations may have been frostier than a winter morning in Lapland after Intel ran into problems with 10-nanometer chips and Nokia consequently suffered. The Finnish vendor has retained Intel as a 5G supplier, using its chips for some radio access network (RAN) functions. But its latest grievance goes to the very heart of Intel's business.

It is, in short, that Intel's general-purpose processors are unsuitable for the most demanding RAN software – a chunk of code that processes signals, commonly referred to as Layer 1 in the industry. Diverging from Ericsson, its Swedish rival, Nokia has decided against incorporating those x86 chips into its cloud RAN portfolio for Layer 1 functions. At an update with reporters in Oulu last week, Nokia executives slammed forthcoming Intel products as the worst example of supplier dependency and lock-in. They are now talking about cloud RAN products that would exclude Intel altogether.

In cloud RAN, the basic idea is to make as much use as possible of general-purpose hardware and common software tools for maximum flexibility. With this sort of infrastructure, the RAN could theoretically share resources with other telco and IT workloads. It could rest on a management platform that stretches horizontally across this whole estate, hosted in private telco or public cloud facilities.

This is bread and butter for Intel. Based on its own x86 architecture, Intel's general-purpose chips accounted for about 71% of the central processing units (CPUs) shipped to data centers last year, according to Counterpoint Research. Intel is pitching these same Xeon-branded processors for use in cloud RAN. Yet Nokia remains unimpressed.

Diminishing returns

One of the company's chief objections is based on a view that Moore's Law is effectively dead. Named after former Intel CEO Gordon Moore, this posits that the number of transistors in an integrated circuit will double every two years, boosting the capacity and capability of a CPU. "Nowadays, with every generation of pure general-purpose computing, there is almost no efficiency improvement," said Brian Cho, Nokia's chief technology officer for Europe and a former Intel engineer. "Moore's Law ended, and this opens the door for a so-called special-purpose processor."

Cho's broad thesis is that general-purpose silicon is still valid when the workload it supports is unknown or unpredictable and some flexibility is required. But this is not the case for an array of workloads, explaining the widespread use of specialist silicon in the cloud. Examples include Nvidia's graphical processing units (GPUs) for artificial intelligence (AI) and games, the data processing units (DPUs) that handle TCP/IP and the tensor processing units (TPUs) developed by Google for AI inference.

Layer 1 falls into the same category, as far as Cho is concerned. Rather than putting this software on Intel CPUs, Nokia says it has worked with Marvell Technology, a US chipmaker, to co-develop Layer 1 silicon deployable in either purpose-built or cloud RAN infrastructure. The result can be thought of as a radio processing unit, Cho told reporters invited on a tour of the Oulu facilities last week. It beats Intel's technology on performance and efficiency hands down, Nokia claims.

Yet the industry still awaits real-world evidence, with these chips not expected to show up in commercial cloud RAN deployments until late next year. In the meantime, detractors led by Intel insist that what Nokia and Marvell have done is not genuinely cloud-native. Custom silicon needs proprietary software and people to manage it, they argue. That software cannot be lifted and shifted to other hardware. Resource sharing is more difficult. This Layer 1, effectively, is not a part of the horizontal cloud platform.

Even Nokia seems prepared to concede there have been trade-offs. Marvell contributes some of the software used with its silicon, and this at the very least is not portable to another chip. Ericsson, by contrast, appears less reliant on Intel in this area. It claims to have built its Layer 1 from the ground up, avoiding Intel's FlexRAN reference design. Only a function known as forward error correction (FEC) comes from Intel.

While resource-hungry, FEC is not where the innovation happens, according to the Swedish company. Thanks to BBDev, a standardized interface between this "accelerated" function and the rest of Layer 1, Ericsson says it has been able to use the code it originally wrote for Intel chips on a CPU made by AMD, Intel's main competitor. While AMD works with the same x86 architecture as Intel, Ericsson hopes to demonstrate portability with CPUs based on the rival Arm setup in the future. This is what openness is all about, Ericsson executives pointedly note.

More Arm wrestling

But Cho thinks Nokia would have a much bigger problem, and far less flexibility, if it had chosen Intel for Layer 1. In Sapphire Rapids EE (SPR-EE), a version of its CPU aimed at RAN operators, Intel has integrated the FEC-handling accelerator with the CPU on the same die. The forthcoming Granite Rapids product seems to go even further by including fronthaul silicon, providing connectivity to radios, on that die. To Cho, it looks increasingly like a system-on-a-chip rather than a CPU. "This is deviating from the mainstream Xeon processor," he said.

Slideware shown to reporters featured scathing language about Intel and "lookaside," the technique it favors that keeps most functions on the CPU. The "inline" alternative that cuts Layer 1 out of the CPU is typically provided on a separate PCIe card, allowing Nokia to "disaggregate" Layer 1 from other network functions. Intel's push in the opposite direction to integrate these various parts leaves customers facing "considerable lock-in," said Nokia.

"SPR-EE is not a general CPU but a purpose-built one," wrote Nokia. The "locked-in portion is the entire server." A customer cannot tinker separately with Layer 1 but must adapt the whole server. Adding Layer 1 capacity means buying an extra server, even if nothing else changes. Customers are also stuck within the x86 ecosystem, says Nokia.

This is potentially significant. Arm-based alternatives are now appearing, and they seem to be a priority for some telcos. Instead of buying their own server equipment, operators may choose to pay for capacity on hyperscaler platforms featuring Arm-based silicon. Nokia already says it can split its RAN software between the Layer 1 silicon co-developed with Marvell and the Arm-based Graviton CPUs designed by Amazon Web Services (AWS). Microsoft's recent launch of an Arm-based CPU branded Cobalt could represent another option.

These hyperscaler moves have buoyed Arm, still better known for its role in smartphones, as a CPU rival to x86. Graviton alone had a 3.16% share of the market for server CPUs last year, according to Counterpoint, up from 1.82% the year before, and Arm-based CPU shipments generated annual revenues of more than $1 billion in 2022 for the first time ever. Given activities by Arm licensees including Nvidia and Oracle-backed Ampere Computing, there is likely to be another market share gain this year.

That appeals to Nokia. While the code produced for Layer 1 is tightly coupled with Marvell's silicon, there does not appear to be the same hardware and software dependency in the upper layers of the RAN. "It is a little bit easier the higher you go," said Veijo Kontas, the head of Nokia's system-on-a-chip division. "There is more general-purpose software, and it is running on Linux." Today, Nokia uses Intel for Layers 2 and 3 in its purpose-built kit. In the future, it should theoretically be able to carry its software to Arm-based rivals without making dramatic changes.

The curious case of Ericsson

Meanwhile, speculation about Ericsson's status and motives extends beyond Nokia. The Finnish vendor champions Marvell partly because it was already used in purpose-built kit. This allows Nokia to ensure there is "feature parity" in cloud RAN and to have only one development track. The implication is that Ericsson has two: one for cloud RAN and one for its traditional gear.

Just last week, the Swedish company was showing off new baseband products for purpose-built 5G. Developed by its in-house Ericsson Silicon unit, they feature custom-made, system-on-a-chip technology, said Ericsson in a statement. Intel is identified as the manufacturer, but these processors are not described in the language of its general-purpose Xeon range.

A hyperlink to a section about Ericsson Silicon also compares that unit's output with other chips. And a chart on power consumption provides an unflattering assessment of general-purpose goods. The seeming contradiction is Ericsson's argument elsewhere that lookaside – heavily reliant on general-purpose processors – is more power efficient than inline, where custom-made silicon is prominent. This claim featured in a white paper that Ericsson co-authored with Verizon last year.

Some of Nokia's criticisms about Ericsson's cloud RAN strategy seem outdated and unfair. One is that Ericsson works only through its own cloud platform, dubbed CNIS. Yet Ericsson has in recent months advertised a RAN tie-up with Google Cloud. Last week it boasted compatibility with Red Hat, the IBM-owned company that was named as Nokia's primary cloud partner in the summer.

Suspicion lingers in the industry that Ericsson is not serious about cloud RAN and its close relative open RAN, conceived to aid interoperability between suppliers. Indeed, there is a view that Ericsson may even be setting these technologies up to fail. But it is certainly not the only RAN vendor that backs Intel, which still enjoys a considerable lead in the small cloud RAN market. If Nokia has its way, that will quickly start to change.