Sachin Katti is still named as an associate professor at Stanford University, and he typically exudes the air of a mild-mannered academic. But in his industry role as chief technology officer of Intel's network and edge group, he occasionally breathes fire.
Katti dismisses the latest merchant silicon for radio access networks (RANs) from rivals like Marvell and Qualcomm as virtualization fakery. His message for Ericsson, an Intel partner? Your custom silicon will eventually be caught and outrun by Intel's x86-based chips, the general-purpose processors underpinning a business worth $79 billion in annual sales.
It is a counterintuitive message. Chips are much like suits. No mass-produced, off-the-shelf outfit accentuates Daniel Craig's gym physique as well as something tailormade, or allows the last actor who played James Bond to move quite so efficiently in tuxedoed pursuit of his enemies. By the same token, an all-purpose chip should never be able to perform as well at a specific task as one designed for that task.
"The way technology is evolving, the gap will reduce, but there will always be a gap between purpose-built and general purpose," said Fredrik Jejdling, the head of Ericsson's networks business.
Figure 1: Vodafone's Yago Tenorio is excited by the range of Intel alternatives.
(Source: Iain Morris/Light Reading)
Yet others outside Intel also foresee a tough future for Ericsson. They include Yago Tenorio, the networks architecture director for Vodafone, one of the Swedish vendor's biggest customers, and the current chair of the Telecom Infra Project (TIP), a Facebook-conceived effort to spur network innovation and lower costs.
"We think the performance of the resulting system combining x86 with acceleration is going to be on a par with or beyond the performance of traditional architecture less than two years from now," Tenorio told a TIP conference in Madrid last month.
To the uninitiated, it must sound as if he is firmly aligned with Intel. But he is not. The "acceleration" Tenorio highlights is about offloading resource-intensive network functions onto other hardware platforms – so-called accelerators – and thereby relieving pressure on the suboptimal x86 used in virtualized systems. It forms part of a broader shift from x86 central processing units (CPUs) to more efficient silicon platforms for storage, security and network functions. That partly explains why Intel's share of the market for data center servers fell from 98% in 2017 to 77% last year, according to Omdia (a sister company to Light Reading).
And acceleration is why smoke billows from Katti's nostrils.
Arm and dangerous
Katti's rivals have one thing in common. Broadcom, Marvell, Nvidia, Qualcomm and Xilinx (a part of AMD) all rely on chip blueprints drawn up by Arm. Based in the UK, and owned since 2016 by Japan's SoftBank, the chip designer generated less than $2.7 billion in sales last year, making it a midget next to Intel. But Arm's simpler designs score top marks on power efficiency, which is of huge importance in the telecom sector, and its ecosystem is large.
Together, those various licensees reported revenues of $119.5 billion in their last fiscal year. Some have already combined their own special sauce with Arm's blueprints in traditional RAN architecture, customized for the needs of a big vendor like Nokia. As Intel's rivals target the same approach in merchant silicon, sold off-the-shelf to anyone who wants it, they are threatening its foothold in open RAN, the nascent market for buying products from multiple suppliers as opposed to a single Ericsson-like system.
Figure 2: Sales of US chipmakers in most recent fiscal year ($M) (Source: Securities and Exchange Commission)
Inline, the acceleration technique preferred by Arm's licensees, essentially diverts the resource-intensive Layer 1 traffic away from the x86 chip, like a bypass around a busy town center. Traffic is instead routed via a separate accelerator card featuring Arm cores and customized silicon.
It is this approach that seems to have caught the interest of Vodafone, which is now working to integrate Marvell's inline accelerators with radio technology from Samsung, a South Korean vendor replacing Vodafone's Huawei network in the UK. Marvell's introduction will mean that project has less need for Intel's latest family of high-performing x86 chips, named Xeon.
Intel has retaliated. Its initial salvo involved an acceleration technique called lookaside. To stick with the road analogy, this is more like a partial bypass, where some traffic still flows to and from the town center. Lookaside required a separate accelerator card for a key Layer 1 function known as forward error correction, but it continued to rely heavily on the x86 chip, protecting Intel's main business. This year, it has been succeeded by what Intel calls an integrated accelerator, which ditches the card and packages various components into the same system-on-a-chip. "Everything is integrated, and you no longer need a separate accelerator card," said Katti.
That is a key marketing point for Intel. Its jettisoning of cards has evidently emboldened the company to criticize rivals still using them with inline acceleration.
"If you do that, that part of the workload is like a physical network function," said Katti. "It is no longer a VNF [virtual network function] or a CNF [cloud network function] because it is literally a physical appliance doing just Layer 1. You can't get it to do anything else, and you can't control it with Kubernetes," he added, referring to the well-known container management platform.
Want to know more about 5G? Check out our dedicated 5G content channel here on Light Reading.
Katti is not completely wrong, his rivals concede. Their general response seems to be that no operator would need this virtualization flexibility in the RAN.
"No one is talking about balancing baseband loads with other applications," said Joel Brand, Marvell's senior director of product marketing. Even if they were, few operators are centralizing baseband resources where sharing workloads would be feasible, he added.
Gerardo Giaretta, the head of Qualcomm's 5G infrastructure business, is in broad agreement. "If you have free CPU cores, it is unclear what you would do with them," he said. "The other thing is that most deployments are at the cell site and there is nothing else you can do with them."
But Giaretta disputes the notion that using Qualcomm's accelerators makes building a cloud RAN or using Kubernetes more difficult. As recently as September, Qualcomm began showing off a setup with Japan's NTT DoCoMo based on Red Hat's OpenShift technology, a managed Kubernetes platform. "You can still use a containerized framework even for this accelerator card," he said.
R&D-off
To others, this contretemps about virtualization is less important than Intel's boasts about the competitiveness of its latest technology. Requesting anonymity, one high-up source in the telco community describes the integrated accelerator as a collection of "chiplets" still tied to the Xeon platform and ultimately stuck with the same problems.
Yet Intel has never sounded so bullish. "General-purpose technology will eventually have so much investment in it that it will outpace custom silicon," said Katti.
His basic rationale is that a RAN market generating about $40 billion a year in sales is simply too small to compete. "Every chip is a pretty large investment of nine figures, and the RAN market is not big enough for someone to spend that amount every year keeping up with the process node improvements," Katti said.
Figure 3: R&D spending by US chipmakers in most recent fiscal year ($M) (Source: Securities and Exchange Commission)
Intel does have a huge and growing research-and-development (R&D) budget, having spent $15.2 billion in this area last year, a 19% increase on the 2020 figure. That dwarfs the roughly $8.1 billion that Ericsson and Nokia spent in 2021, and much of that did not go into silicon-related activities. But the US licensees of Arm challenging Intel in the RAN collectively spent $22.6 billion. And while the RAN accounts for a small share of their business, RAN accelerators can feasibly piggyback on R&D in other business areas.
In the RAN, moreover, Intel's Xeon processors are more about what happens inside distributed units, the server boxes where most baseband processing is done. They do not have much bearing on the radio units, the other critical element. "If you look at the way we design massive MIMO, where the radio unit weighs 12 kilograms, that is because we build a lot of integration into the chipset at the radio level as well," said Ericsson's Jejdling, referring to an advanced type of 5G technology.
The bigger risk for Ericsson and Nokia, arguably, is that a market for Arm-based merchant silicon takes off, threatening their own silicon investments. Marvell already figures prominently as a supplier of baseband chips to Nokia for traditional 5G architecture, under the Finnish vendor's ReefShark brand. What if other RAN software companies could easily take advantage of the same silicon? "Nokia insists ReefShark is fully customized for them, but I suspect it is just merchant silicon and only the software is customized," said Tenorio.
Such considerations might explain why Ericsson appears to be in the Intel camp. Acknowledging the desire of some customers to use virtual RAN technology, it now allows its baseband software to be run on Xeon processors and Intel accelerators as an alternative to its own customized silicon. Take-up of the former squeezes the latter, Jejdling admits. But a multiplicity of Arm licensees and their software partners could appear far more dangerous.
Figure 4: Ericsson's Fredrik Jejdling says silicon investments have produced more lightweight radios.
(Source: Ericsson, YouTube)
Meanwhile, the virtual RAN market is small, Intel-dominated and currently fueled by a Verizon rollout in the US. "There might be other small deployments, but they are not making a big difference," said Remy Pascal, an analyst with Omdia.
Nor is Ericsson the only Intel enthusiast. The chipmaker appears to have a major ally in Tareq Amin, after helping the boss of Rakuten Mobile to launch one of the world's first virtualized mobile networks in Japan more than two years ago. During an interview he recently gave to The Mobile Network, Amin derided the use of "additive acceleration cards," appearing to echo Katti's criticisms. An executive at a RAN software company, who did not wish to be named, described them as a costly complication.
Inflexible RAN
Both a blessing and a curse for Intel is a tool it calls FlexRAN. It has existed for many years as a kind of reference guide for software companies developing virtual RAN technology, provided free-of-charge by Intel. An ecosystem has grown up around FlexRAN, sucking other companies into its orbit. But software written on top of it is not portable to other accelerators, according to Vodafone's Tenorio.
"You have to design your software in a particular way using the instruction set FlexRAN gives you," he said. "Once coded for FlexRAN, it is not portable for another accelerator – you can only use Intel from that point on."
Lack of standardization at this level is not restricted to Intel, though. The instruction sets from its rivals also tie software companies to specific accelerators, said Tenorio, calling for a new standardization push. The O-RAN Alliance, the group behind open RAN specifications, is trying to narrow the portability gap via an accelerator adaptation layer and the use of standard APIs (application programming interfaces).
Working Group 6, the part of the O-RAN Alliance charged with that effort, has already had success at opening the interfaces between Layer 1 and Layer 2, allowing Intel CPUs to coexist more easily with third-party accelerators. But the latest task looks technically difficult. "What we may need to do next is some harmonization of instruction sets to make software more portable," said Tenorio.
Figure 5: Tough year for semiconductor stocks (Source: Google Finance)
The O-RAN Alliance's work faces opponents, too, with Ericsson understood to be among them. While smaller software players might welcome less effort per accelerator, too much standardizing here could make differentiation of software products that much harder. If most of the R&D spending by Ericsson and Nokia goes into software customization, they would both have a lot to lose.
This would all make Katti's forecast rather troubling for the Nordic kit vendors if they took it seriously. "When merchant silicon is better than custom silicon, does it make sense to build custom silicon or focus resources on the software that is on top of it, where they deliver the most differentiation?" he said. "If anything, I think it will allow them to focus resources on the thing that differentiates them the most."
Tenorio dreams of a vibrant and competitive open RAN market for merchant silicon, where Vodafone can choose from numerous blends of RAN software and accelerators. If he is right, then Ericsson's wariness about open RAN and Nokia's hokey-cokey with it are understandable. And Intel's fighting talk makes sense. Virtual RAN might be a small outpost in the empire. But losing it would be a bad sign.
Related posts:
— Iain Morris, International Editor, Light Reading