Ericsson and Intel are at odds on chip vision

When AT&T starts to build its "open" radio access network (RAN), after striking a $14 billion deal with Ericsson, it won't be using Intel's chips for the most ravenous software. "To start, we will leverage the strength of Ericsson's silicon roadmap to introduce open radios onto our network," an AT&T spokesperson previously said by email. "As silicon solutions mature, we will introduce additional open architecture solutions based on the Intel roadmap."

It's confirmation this AT&T network will not, at first, be a "cloud" RAN, despite the naming of both Intel and server partner Dell in AT&T's official release. Networks are typically purpose-built, fusing custom chips with proprietary software at the outset. In a cloud RAN, the idea is to write code in a common language and deploy it on the same general-purpose processors that can support multifarious other workloads. Intel's Xeon processors, installed in servers made by Dell, are central to Ericsson's approach.

But as far as the Swedish kit maker is concerned, those Intel chips will never beat its own silicon in one critical area, despite recent gains. "Purpose-built hardware is currently more energy efficient for the RAN application than general-purpose hardware, although with the planned improvements in future processors the gap will close," said Michael Begley, the head of Ericsson's RAN compute product line, by email. "However, given its optimization point, purpose-built hardware will continue to be the most energy-efficient and compact hardware for radio site deployments going forward."

Related:Partner Report - 5G in South-Eastern Asia - Essential insights into the future of 5G across the region.

In other words, the gap will never entirely close, implying AT&T will be worse off with any cloud RAN it eventually builds. Why do it, then? Begley's rationale is that cloud RAN has other attractions that might be more important in some scenarios. "The operator wants to pool edge hardware resources, repurpose them according to need and be able to fully utilize all the resources as and when needed," he said.

Difference of opinions

Yet the performance of general-purpose processors is a topic on which Ericsson and Intel seemingly disagree. "General-purpose technology will eventually have so much investment in it that it will outpace custom silicon," said Sachin Katti, the general manager of Intel's network and edge group, during a previous conversation. Intel is pumping billions into the miniaturization of chip transistors so that more can be crammed into a single processor. The RAN market, in Katti's view, is not big enough for specialists to keep up.

If he is right, then Xeon processors will eventually catch and overtake Ericsson's application-specific integrated circuits (ASICs) and system-on-a-chip (SoC) technology. In that situation, Ericsson's continued investment in baseband silicon would look wasteful. The more sensible strategy? Adopt general-purpose processors across the entire RAN portfolio and double down on software development instead.

Katti has previously urged Ericsson to do just this, arguing in late 2022 that software is "where they deliver the most differentiation." It would naturally suit Intel to see one of its biggest customers in the telecom sector swing wholeheartedly behind general-purpose processors, today's market for which is still dominated by the giant US chipmaker. Yet even outside Intel, there is a rumor that Ericsson is planning such a move.

It has, no doubt, acquired impetus since last July, when Ericsson named Intel as the company that will make its 5G network chips in the future. Under the arrangement publicized then, Ericsson will rely on Intel's "18A process" – where transistors measure less than 2 nanometers, or billionths of a meter – to produce its latest chips. Was this proof of an Ericsson retreat from in-house silicon development?

Absolutely not, according to Begley. "Investing in Ericsson Silicon will remain at the heart of our strategy as it delivers unique customer value," he said. "Ericsson will continue to develop its custom 5G SoCs for all of its purpose-built hardware going forward." Intel's role in this area is as a manufacturing partner only, Begley explained. Ericsson, of course, has no large semiconductor manufacturing facilities of its own.

Customizing the cloud

But the intrigue does not end there. A few well-placed sources believe the relationship between Intel and Ericsson differs little from a tie-up between Marvell Technology and Nokia. The Finnish rival to Ericsson has been combining its intellectual property on the radio side with Marvell's silicon expertise to produce chips for Layer 1 RAN functions, those hungriest for resources. Nokia's strategy is to use Marvell's Layer 1 chips across the entire RAN portfolio, with general-purpose processors retained only for less demanding functions (Layers 2 and 3) in cloud RAN deployments.

Each camp now has similar criticisms of the other. Those on the side of Ericsson and Intel have long insisted that Nokia and Marvell are effectively excluding Layer 1 from cloudification. Even Nokia has acknowledged that Layer 1 code would need rewriting if it switched from Marvell to another chip. But its supporters would argue that Ericsson looks equally dependent on Intel. Some of the central processing units (CPUs) intended for cloud RAN are "actually not general-purpose CPUs," said Nokia in an October white paper. "They are essentially cloud RAN-specific CPUs."

Customization does seem to be a growing feature of Intel's cloud RAN portfolio. Sapphire Rapids EE is not just a standalone CPU but also comes with a fully integrated, Intel-specific hardware "accelerator" to handle forward error correction, an especially demanding bit of Layer 1. In Granite Rapids, a future release, some of the channel-estimation processing in massive MIMO, an advanced 5G technology, would also move from the CPU to this accelerator, according to a reliable source. Fronthaul connectivity is included on the same die, as well.

The casualty in all this seems to be any notion that common, off-the-shelf hardware running more easily programmable software can meet the full needs of the telecom network. That should probably not be a surprise. So-called "accelerated computing" is now a widespread feature of the cloud industry, with graphical processing units and more customized chips used in place of CPUs for numerous workloads. But it's a story that doesn't please everyone.