Tareq Amin has never been afraid to cut his vendors loose if he thinks they are wrong for the job. Since joining Japan's Rakuten in 2019, the man now in charge of its telecom subsidiaries has switched from Cisco to NEC in his mobile network core, dumped Red Hat as a cloud platform after buying a small rival called Robin.io, and bid sayonara to Nokia for radios, preferring NEC and others with the leap from 4G to 5G technology.

But one of the oldest changes took place at Altiostar, the US software company that Amin eventually bought to power his mobile network. Long before that deal happened, Altiostar had been relying on chips developed by Texas Instruments, based on blueprints drawn up by UK chip designer Arm. One of the founding fathers of the semiconductor industry, Texas Instruments had its origins in designing precision-guided missiles for the US military. But a small part of its own twenty-first-century civilian business took a direct hit when Amin, Altiostar's only big customer at the time, pushed his supplier to use Intel instead.

"I could have run virtual software on an SoC [system-on-a-chip] that is Arm, but I also wanted it to integrate properly with my cloud environment," said Amin during a recent interview with Light Reading. By convincing Altiostar to port its code to Intel's general-purpose processors (GPPs), Amin could stand his radio access network (RAN) software on the same chip foundations supporting Rakuten's other network and IT workloads.

Figure 1: Rakuten's Tareq Amin showing off a radio kit at Mobile World Congress 2022.
(Source: Iain Morris/Light Reading)

His remark exposes one of the main problems in this market – the lack of viable GPP alternatives to Intel. Numbers crunched by Counterpoint Research show that Intel's x86 chips were used as central processing units (CPUs) in 70% of all notebook computers shipped last year. In the market for server CPUs, Intel's market share was similar at 71%.

The only other sizeable player in both markets was AMD, with about 18% of notebook computers and 20% of servers, and AMD uses the same x86 architecture as Intel. Licensees of Arm, the main alternative system, collectively accounted for less than 13% of notebook computers and a measly 5% of the market for servers in 2022.

In the virtual RAN market, where network software is paired with GPPs, Intel's dominance is even more pronounced. During a press briefing at the start of the year, Sachin Katti, the head of Intel's network and edge group, claimed Intel was present in about 99% of the world's networks. Nobody worries much because virtual RAN itself accounts for only a tiny percentage of the total RAN market. But the analyst consensus is that networks will become progressively more virtualized this decade. A RAN sector dominated by Intel would be at odds with the industry's push for greater supplier choice.

Figure 2: Notebook shipment share by CPU type (Source: Counterpoint Research)

RISC factors

Efforts to foment Arm-based virtual RAN competition to Intel have so far largely focused on Layer 1, the part of the RAN software stack responsible for baseband, the processing of radio signals. In a virtual RAN, most of this happens in a box called the distributed unit (DU), typically found at the base of a mast or in nearby facilities. Through a technique called inline acceleration, an operator could move the Layer 1 processing onto separate Arm-based silicon, usually delivered via an accelerator card that slots into a DU server. Compatibility with the server kit is guaranteed thanks to a standard interface called PCIe.

Telcos are exploring various options for accelerators because x86 processors on their own do a relatively poor job at Layer 1, guzzling far more energy than the customized chips included in traditional RAN equipment. Marvell and Qualcomm are among the chip companies to have started offering PCIe cards in response to what they see as telco demand.

But doubts have been raised. Geetha Ram was surprised by the high price being charged for a PCIe card she examined in her role as the head of telco compute for HPE, one of the world's biggest server makers. Rakuten's Amin has ruled out ever using them. "It is extremely inefficient," he said. "If I can buy one CPU that does the job of this accelerator card that sits somewhere else, it is cheaper."

Even if a telco disagrees about the expense, a PCIe card still needs a server with a CPU, which probably comes from Intel. The company has responded to the launch of inline accelerators for Layer 1 by investing in an alternative technique called lookaside. It can be integrated with the CPU, cutting the need for PCIe cards, and is championed by Ericsson and Verizon as a more energy-efficient option. The danger is that it reinforces Intel's Layer 1 dominance. Regardless, take-up of PCIe cards would not end reliance on Intel or x86 for other RAN functions.

That can only realistically happen with the emergence of CPUs developed by rivals using other chip architecture. HPE is backing Arm to be that alternative. Having already developed some Arm-based servers for core networks, it is currently working on potential RAN servers in partnership with an Arm licensee called Ampere Computing, founded in 2017 by Renée James, a former president of Intel.

Ram's goal is to build something with fewer cores, a term for the smaller processors that make up the overall CPU. Historical comparisons of Arm and x86 have noted that Arm runs at a slower clock speed, a measure of its processing power, and has therefore needed a higher number of cores. But a 128-core Ampere CPU would be "overkill" in the RAN, said Ram. "We are trying to see what we can do at a lower core count of Arm."

Figure 3: An Apple R&D facility in Munich, Germany, where silicon is designed.
(Source: Apple)

She was impressed, however, when assessing Ampere's 128-core CPUs next to 32-core CPUs bought from Intel. "You get anywhere from two to four times in performance and the power is about the same," she said. "That is just a fact." In Ram's view, those advantages stem from the underlying RISC (reduced instruction set computing) architecture fundamental to Arm (whose very name stands for Advanced RISC Machines). The older CISC (complex instruction set computing) architecture integral to x86 looks cumbersome alongside it. At one time, even Intel considered switching.

There are other reasons to be optimistic about Arm. Two years ago, it launched a new high-performance core branded V1 that addresses some of its earlier shortcomings. Unlike Intel, moreover, its licensees are entirely fabless (meaning they do not manufacture their own chips) and can therefore take advantage of the most advanced process nodes, typically measured in nanometers (where smaller is better), without handing money to a rival. Ampere uses TSMC, the world's most advanced foundry. Marvell has commercial RAN chips based on five-nanometer nodes, currently beyond Intel's ability to produce.

Juicier targets

But if Arm is so good, why has it not had more success outside the smartphone market it dominates? Largely, it seems, because the PC and server markets, after previously tying themselves to x86, have proven stubbornly resistant to any disruptors. The optimization of Microsoft Windows software for x86 chips has built a wall around PCs. Switching the underlying workhorse technology in data centers is likely to be harder than just upgrading it.

Even so, Intel's share of key markets is suddenly dropping faster than ever, and Arm's is rising. In 2021, Counterpoint reckons, its share of the notebook computer market soared from 2.7% the year before to a meatier 10.7% after Apple ditched Intel in favor of its own M1 processors, based on Arm blueprints. Most reviews have been ecstatic.

Arm's progress in the data center has been much slower. Most of the gains at Intel's expense have gone to AMD, the only other big supplier of x86 chips. But one recent trend has been the development of Arm-based CPUs by Internet giants. Amazon now has its own Graviton line of processors, for instance. Its share of the server market grew from about 1.8% in 2021 to roughly 3.2% last year, according to Counterpoint.

Rakuten's Amin craves a chip like the M1, which integrates accelerators and other important features on the same die. But Apple does not sell merchant silicon in the RAN market, and neither does Amazon. For companies that aspire to challenge Intel, the computer and server markets are much juicier targets. Even after its tumbles, Intel made $50.9 billion in those markets last year, while the entire market for RAN equipment was worth only $45.3 billion, according to Omdia (a Light Reading sister company), with semiconductors accounting for a fraction of that amount.

Figure 4: Sales at major Intel units ($B) (Source: Intel)

Today, Intel's biggest challenge arguably comes from Qualcomm. The Arm licensee, known chiefly for its importance in the smartphone sector, is working on a chip to power notebooks. Unlike the M1 processor (and M2 successor) that Apple designed for its own Macs, this would run on Windows PCs. "If this new family of chips lives up to expectations, it will represent a dire threat to all Intel shipments into the laptop market," said Richard Windsor, an analyst with Radio Free Mobile, in a recent blog.

What's interesting is that Qualcomm also makes silicon for the RAN. If its PC ambitions were realized, a more aggressive push into the telecom network fueled by that success would be no radical departure, unlike a RAN silicon move by Apple or Amazon. Others cite Nvidia, a maker of graphical processing units (GPUs) preferred to CPUs for some data-center workloads, as one to watch, while Amin hails the brilliance of Marvell's five-nanometer designs.

Joel Brand, a senior director of product marketing at Marvell, envisages a future in which virtual RAN servers look more like the appliances – essentially, the proprietary kit – used in a traditional RAN. "The server that HPE or Dell is selling to telcos is already optimized," he said during a previous interview. "It is already a unique server, and someone will build them a unique server that will be Graviton or Google Cloud or Microsoft."

Yet the downside of having Intel as a dominant force in GPPs must be weighed against the benefits of using a single technology to support a vast array of network and IT systems. Relying on abstraction layers built by cloud companies to smooth over the differences between underlying instruction sets and tools might simply create new choke points. With the displacement of data-center CPUs by more customized chips, such as GPUs, a vision of being able to run any software on any silicon will strike many as unrealistic.

Much harder to dispute is that Arm will figure more prominently in the virtual RAN market later this decade. "It will go down the route of something more optimized for the needs of the workload," predicted Brand. "I think it is going to be with Arm CPUs just because I see all the cloud guys adopting Arm." The big uncertainty is just how far Arm can advance. With its resources and installed base, Intel is not about to vanish. But its recent difficulties do not inspire confidence.

Related posts:

— Iain Morris, International Editor, Light Reading