Intel inside, possibly the best-known marketing slogan in the technology sector, neatly encapsulates the giant chipmaker's ubiquity. Trying to find a PC that does not feature Intel inside is nearly impossible. Estimated at between 90% and 95%, its share of the market for processors used in data centers easily makes it the dominant player. Many would call it a monopoly.
No one would apply that description to Intel in mobile. The chips that power most of today's radio access network (RAN) systems come from a strong line-up of semiconductor specialists, including Broadcom, Marvell, Xilinx and Huawei-owned HiSilicon. Yet Intel is increasingly visible in that mix. What it calls its network platforms group made just $1 billion in revenues in 2014, a tiny fraction of the $56 billion Intel generated in total sales. By 2020, it had grown into a $6 billion business.
Much of that growth has come from Intel's success in the core network and enterprise markets. Virtualization, allowing companies to run software on Intel's general-purpose processors, has taken hold in these sectors. Back in 2013, there were almost no virtual network servers. Last year, they accounted for about half of all core network deployments, and they are projected to reach 80% by 2024, Intel reckons.
The chipmaker's advance into the mobile access market has been even more dramatic. From having almost zero presence in basestation silicon just a few years ago, Intel today boasts a 40% market share. In this case, virtualization is not responsible. Codenamed Snow Ridge, its latest Atom P5900 processor is sold mainly to the makers of traditional mobile networks. Ericsson, Nokia and ZTE are among Intel's customers.
But virtualization is also nibbling into the RAN. Service providers have latched onto a concept called open RAN, believing it will loosen the tight market grip of Ericsson, Huawei and Nokia. It promises new interfaces that would let an operator mix products from different suppliers at the same site, instead of taking all the parts from one of those giant vendors. Because open RAN and virtualization go together like kama and sutra, it would also rely heavily on general-purpose processors. And that puts Intel very firmly on the inside.
How big open RAN will eventually become is still up for debate. By the mid-2020s, analyst firms Dell'Oro and Omdia (a sister company to Light Reading) both think open RAN will account for roughly 10% of a RAN market worth about $30 billion annually. But many service providers hope it will be the preferred way of building networks by the late 2020s. France's Orange, for instance, wants all new products it buys to be compatible with open RAN starting in 2025. The irony is that a concept supposedly about competition and supplier diversity could end up placing Intel in control.
Intel's move into the RAN market started in 2010, when it began investing in a software stack that eventually became known as FlexRAN. Four years later, it spent $650 million on Axxia, a RAN system-on-a-chip (SoC) business owned by Avago, which had acquired it with an earlier $6.6 billion takeover of a company called LSI. Axxia provided Intel with its entry point into the market for basestation silicon, where it now claims a 40% share.
FlexRAN, however, is all about virtualization. In this scenario, the RAN is broken up into three main blocks. First, there is the radio unit (RU) atop a mast, which receives and handles the radiofrequency signals. Next comes what is known as the distributed unit (DU), a box near the mast site (sometimes at the foot of it) that is mainly responsible for "baseband," the processing of those signals. Further away is the central unit (CU) that looks after the control plane, the network's traffic cop. Software based on FlexRAN would run over Intel's latest x86-based Xeon processors in the DU.
The various functions split between these elements are often described with reference to conceptual layers, or different categories of function. The baseband functions handled by the DU are classed as Layer 1 (sometimes called the physical or PHY layer), although some Layer 1 processing might also move into the RU, especially in high-performance networks. Further up is Layer 2, the data link or MAC (for media access controller) layer, whose functions are divided between the DU and CU. The CU would also be where network layer processing – so-called Layer 3 – takes place.
Outside the pure radiofrequency part of all this, virtualization gives Intel a role across the entire chain. Its opportunity in the CU is a given, says Simon Stanley, a semiconductor expert and principal consultant at Earlswood Marketing, because "that lends itself to putting on servers and virtualizing very easily." But a mixture of the Atom P5900, Xeon and FlexRAN is powering Intel into other parts of the RAN.
Dan Rodriguez, the man in charge of Intel's network platforms group, explains how virtualization is a bridge to new territory. In a traditional RAN, Intel's Atom P5900 would look after the control plane and leave signal processing to an application-specific integrated circuit (ASIC), developed partly by an in-house team at Ericsson or another one of the big RAN vendors. "In the virtual RAN and open RAN world, the control, packet and signal processing are all done on Xeon and that is what FlexRAN enables," he says.
The other guys
The worry for operators is the lack of viable alternatives to this Intel set-up. Competition is expected from the chipmakers that use blueprints supplied by Arm, a UK firm whose application processor designs underpin most of the world's smartphones. Those include some of the firms that Intel has been challenging with the Atom P5900 – the likes of Marvell and Xilinx. Yet just as Intel has largely failed to crack the smartphone sector, so Arm has struggled in the market for server infrastructure.
"Why is Intel popular in this? What else is there?" asks Vinay Ravuri, a former Qualcomm executive who now runs a chips business called EdgeQ. "In general, compared with Intel, Arm has less of an ecosystem because it is fairly new to the server infrastructure game, period. They are much better on the client side."
But Arm and its partners have a lot to lose if the open and virtualized RAN supplants the traditional RAN in the next decade. Dissect an Ericsson or Nokia network this week and you would probably find Arm designs at its heart, says Panch Chandrasekaran, Arm's director of 5G carrier infrastructure. "The odds are that if you open up any of these hardware boxes today, the dominant CPU [central processing unit] architecture is Arm."
A strict examination of Arm alongside Intel is not really comparing "apples with apples," notes Stanley. While Intel designs and makes its own x86 chips, Arm provides architecture and implementation blueprints to companies including Marvell, Qualcomm and Xilinx. These silicon partners would essentially take Arm's designs, add some of their own magic and sell hardware to Ericsson or Nokia. It is this broader market that Intel now threatens.
Whether or not open and virtualized RAN networks take off, even Chandrasekaran is forced to admit that Intel has established an early lead in that particular race. "It is just a matter of the amount of time that has been spent by our friends on the other side so far," he says. "I think it is fair to say that phase one of open RAN deployments will be pretty much x86-based."
For anyone concerned about competition and choice, the worst-case scenario is that a RAN oligopoly comprising Ericsson, Huawei and Nokia gives way to an open RAN Intel monopoly. Rodriguez rejects this type of assessment. In the traditional RAN, the ASICs, segment-specific SoCs and operating system are all controlled by the same vendor, he points out. Open RAN brings greater choice in areas including server design and network applications development, he says. "Just by default, there is a diverse ecosystem in open RAN and virtual RAN."
But the only company besides Intel that makes x86 processors is AMD, and even it has drawn attention to Intel's dominance. "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," it wrote in its last annual filing with the US Securities and Exchange Commission.
Moreover, while there may be a choice of server makers, one of several software companies that have sprouted around FlexRAN has also acknowledged Intel's status as today's power broker. In comments about open RAN emailed directly to Light Reading, a spokesperson for Parallel Wireless said that "today it is all about Intel and its partners," listing Dell, HPE and Supermicro as examples.
Software, clearly, is a possible choke point for open RAN. Parallel Wireless is one of three US firms that have emerged as the most visible software options for an open RAN network – the others being Altiostar and Mavenir – and all of them have based their systems on FlexRAN, tying them to an x86-based system. For operators keen to make use of Arm's architecture, that is potentially a problem.
An early example is Dish Network, the US satellite firm now in the process of building a greenfield mobile network based on open RAN technology. Dish has already identified Altiostar and Mavenir as software suppliers. But after recently signing a public-cloud deal with AWS, it also highlighted its interest in putting some RAN functions on the Arm-based Graviton2 processors that AWS has developed internally.
It sounds like a head-scratcher for Marc Rouanne, Dish's chief network officer, who is keen to avoid introducing additional software players and complicating the set-up. "You don't want to spread people too thinly by having too many platforms, so we are also working with a number of innovators on having an abstraction layer that could even be distributed or open source or shared and would allow us to abstract the underlying processors," he says. "Then we could port Mavenir and others across different environments."
Following some tests, Rouanne is confident this approach will succeed, but it does not sound ideal. "It is not a big, big thing, but the less your software depends on your underlying hardware, the better it would be," he says. For observers, the question may be whether an abstraction layer on which Dish relies heavily becomes a new type of shackle.
Others in the open RAN market sound unimpressed by FlexRAN. Its critics include Slawomir Pietrzyk, the CEO of a Polish software startup called IS-Wireless. "We think that Intel is working against the core idea of the whole movement, which is being open," he tells Light Reading. "This was our thinking and customers confirming they don't want yet another vendor lock-in."
But Intel's monopolization of the open RAN space so far does not come as any surprise to Stanley. "Anyone doing software always does Intel first because that addresses 95% of today's market," he says. "It is a self-fulfilling prophecy, to some extent, because if 95% of the software is already on Intel it discourages people, unless they have good reasons, to go to Arm. This has been Arm's challenge all along."
Arm and the RAN
Arm is certainly not a lost cause, though. Partners including Marvell and Qualcomm have touted forthcoming open RAN products based on Arm architecture. Working with Benetel, an Irish maker of radio units, and Analog Devices, a radiofrequency specialist, Marvell is promising a full set of RU, DU and CU products, including reference software for Layer 1. Qualcomm, seemingly, is restricting itself to the DU and RU.
This opens up the possibility of coexistence with Intel in the same RAN. Qualcomm's rationale seems to be that it can challenge Intel in baseband but avoid it in higher layers and the CU. EdgeQ, similarly, is developing chips for Layer 1 based on RISC-V, an open source architecture that seems to be the only feasible alternative to x86 and Arm.
"You can put in a server platform and then plug in an accelerator that has got a Qualcomm or Xilinx or Marvell processor on it, in which case you have then got a mix," says Stanley. "You have got Arm running the low-level processing and Intel running the higher-layer processing and so it is not an either-or situation, really."
There is also optimism that new software options will start to emerge, driven partly by operators. A collaboration between Qualcomm and Vodafone on reference designs for open RAN "could produce a solution," says Stanley. Work carried out by the O-RAN Alliance, the main open RAN specifications group, could throw up further alternatives, he says. "They are doing some implementation examples as well and that is all based on open source software."
EdgeQ's Ravuri hopes software will give his company an advantage over Intel and FlexRAN, too. Customers have already been drawn to the technology because of its "uniquely open software programmability," he said via email. "Our choice of RISC-V allows customers an open platform where they can have customability across all levels of the cellular stack (L1/L2/L3), and also benefit from the community of open tools for development."
Arm, meanwhile, now claims to be working with software vendors in China, Europe, Japan and the US on open RAN. While Chandrasekaran will not be drawn at this stage on specific details, he cites Belgium's Accelleran, as well as AccelerCom and CommAgility in the UK, as companies of interest. The big challenge is identifying firms able to scale up their products. "When you look at it that way, the number of players around the globe trickles down to very few," says Chandrasekaran.
For Intel, the unresolved, multi-billion-dollar question is whether x86 can overcome its power-efficiency disadvantage next to more integrated and customized products. Xeon D, its Layer 1 iteration of the processor, comes with an accelerator designed to help, and Rodriguez says customer activity shows operators are "confident" Intel can deliver. "We are going to keep investing in our silicon and keep being more efficient," he says.
But leading semiconductor analysts have not previously been convinced. "No one disputes that you can do Layer 1 with an accelerator using x86 for wireless. The problem is you need a room compared to one box or a rack," said Earl Lum, a semiconductor analyst with EJL Wireless Research, during an earlier conversation with Light Reading.
Nokia, one of Intel's own Snow Ridge customers, has cast similar aspersions on x86, despite its professed support for open RAN. "The reason why established vendors have custom-made system-on-a-chip hardware solutions is to ensure that we meet the targets when it comes to energy efficiency," said Florian Damas, the head of Nokia's regulatory and policy affairs team, during a recent online conference.
The Finnish vendor was obviously burned by its own experience of using field programmable gate arrays (FPGAs) – one of the main options for open RAN hardware acceleration – in its traditional 5G products. It switched to FPGAs supplied by Xilinx after Intel ran into manufacturing problems with customized silicon. The result was a rise in costs and a loss of product competitiveness, which Nokia is still addressing.
Xilinx insists the bad press about FPGAs is increasingly out of date and that capabilities have improved. "I think the people that are raising the flag from the vendors are probably feeling the threat from those types of technology," said Gilles Garcia, the senior director of Xilinx's wired and wireless group, during an interview with Light Reading in April. He declined to comment specifically on Nokia's difficulties. "The only thing we can say is that Nokia was a very large Xilinx customer and today still is a very large Xilinx customer."
Regardless, if operators are determined to use open RAN, Arm and its partners will need to adapt. And its own reputation for energy efficiency provides no certainties. "There is no guarantee that just because you are using Arm you will be lower power," says Stanley. If forthcoming products do not measure up, or are not as cheap as they need to be, then Arm may struggle to build momentum, he says. "That is kind of where we are with Arm at the moment."
While Arm takes stock of the changing market dynamics, it can at least count on further criticism of x86 as a technology unsuitable for all scenarios. "Cisco is a huge company for a reason. They are not going to run a whole router on a regular server," says EdgeQ's Ravuri, alluding to the ongoing need for purpose-built products.
"You cannot have a server solution for everything," insists Chandrasekaran. "Having that power efficiency and performance efficiency is a big, big concern for anyone that is trying to do open RAN. There is no one-size-fits-all kind of thing. Close to the radio tends to be very heterogeneous in nature."
Indeed, for all the interest among service providers, experienced commentators and experts are still not entirely persuaded that open RAN is a revolution in the making. "While the general attitude toward open RAN and the outlook has improved, it still remains somewhat uncertain to what extent open RAN will impact the overall RAN dynamics," said Stefan Pongratz, an analyst at Dell'Oro, in emailed remarks.
Whatever transpires, a market dominated by Intel would somewhat defeat the entire purpose of open RAN. That is probably why Chandrasekaran can afford not to be too worried about Arm's lag. "There is definitely interest from our customer base in making sure there is an ARM-based ecosystem around this," he says. "The worst thing that can happen is for them to go from an Ericsson, Nokia duopoly to a traditional CPU monopoly."
- Open RAN will struggle to fit in, despite Euro telco demands
- Huawei the likely loser in Orange open RAN plan
- Poland's IS-Wireless steps into Europe's open RAN void
- Xilinx and Mavenir hail massive MIMO breakthrough for open RAN
- Open RAN and the mission to crack massive MIMO
— Iain Morris, International Editor, Light Reading