Ericsson CEO Per Narvinger is not for steering the Swedish vendor away from the design of custom RAN silicon.(Source: Ericsson)

Ever since Nokia struck its $1 billion pact with Nvidia around this time last year, industry watchers have wondered if rival Ericsson might do something similar. With Nvidia's investment in Nokia came the revelation of plans by the Finnish vendor to design radio access network (RAN) products based on the huge chipmaker's graphics processing units (GPUs), ordinarily aimed at the data center. Might Ericsson also be persuaded of a GPU's attractions over conventional custom silicon?

The answer has been a polite but firm no, although not without a caveat. For years, Ericsson has designed its own application-specific integrated circuits (ASICs) for Layer 1, the slice of RAN software that covers the most demanding tasks. In the 5G era, the strategy seemed to differentiate it from Nokia, which switched to a partnership with Marvell Technology for custom silicon after initial 5G setbacks with Intel. Under Nokia's latest plans, Nvidia will eventually replace Marvell. Per Narvinger, Ericsson's new boss, sees no reason to follow by substituting GPUs for Ericsson's ASICs.

"We have disaggregated our software and hardware, meaning we have a lot of flexibility on what path to embark on in the future," he told equity analysts on Ericsson's call about third-quarter results last week. "Having said that, right now, we see that it's an extraordinary asset to have your own ASICs. That really makes it possible for us to get our functionality fast. It makes it possible for us to be very cost competitive."

The remarks came after Narvinger was asked if Ericsson had closed the door on putting GPUs into the baseband responsible for RAN compute. Nokia and Nvidia insist GPUs offer advantages over custom silicon, allowing them to swap out normal RAN algorithms for more efficient AI. If evidence of that is provided by ongoing trials with prominent operators like T-Mobile US, Ericsson's defense of custom silicon could start to look seriously misguided.

The spectre of efficiency

But Ericsson's experts clearly doubt Nokia can realize any significant advantages with a GPU RAN. Algorithms and models deployed in a RAN are constrained by factors such as the transmission time interval, which means important calculations cannot take longer than 500 microseconds. Software designed to work in these conditions does not require the processing muscle of a GPU – or so the argument goes.

Nokia claims it can double the spectral efficiency of a network by 2028 with its GPU strategy. The figures published by Ericsson, based on the continued use of custom silicon, are much lower. But various experts including Marcus Weldon, a former Nokia chief technology officer, reckon the huge improvements Nokia has talked about will be limited to a small number of use cases.

The baseline for comparison has also not been specified, but significant gains would naturally be easier to achieve on a badly performing network. "A site already achieving near-theoretical efficiency under classical processing has less headroom for AI-driven improvement than a site operating well below its theoretical ceiling due to interference, channel estimation errors, or suboptimal scheduling," said Chetan Sharma, the founder of Chetan Sharma Consulting, in a recently published white paper.

Importantly, too, Nokia's approach requires operators to install GPUs in their networks, while Ericsson boasts commercially deployed silicon that can already support new AI features such as AI-native link adaptation, a RAN algorithm. "It's been around in the industry for as long as I've been, so more than three decades, and now replacing that with an AI model we get 10% gain in spectral efficiency," said Narvinger. "Today, we have that deployed commercially with three operators, and we have more than 60,000 basestations running this functionality."

Even so, Narvinger told analysts that Ericsson is "not ruling out any other platform," and the Swedish company has partly hedged its bets. The "disaggregation" cited by Narvinger refers to Ericsson's separate virtual or cloud RAN development track, which today runs most Layer 1 functions on Intel's general-purpose central processing units (CPUs). Only a single resource-hungry task – forward error correction (FEC) – is offloaded to a hardware accelerator, which Intel supplies and combines with its latest CPUs.

Ericsson is now confident it could move the software for the CPU-hosted functions to other CPU platforms after small tweaks. Such platforms might include AMD, which uses the same x86 architecture as Intel, or even a processor based on Arm, a rival architecture. During trials, it already claims to have ported software to an Arm-based CPU offered by Nvidia. A GPU, in this kind of scenario, could function as the FEC hardware accelerator while supporting other AI needs.

It's the economy, stupid

But economic factors play a critical role in the RAN choice between custom silicon and general-purpose chips, whether CPUs or GPUs. Ericsson's naysayers insist it will not be able to sustain custom silicon investment in what Narvinger himself describes as a "flat" RAN market. General-purpose processors, by contrast, have a much bigger target market and better returns.

It's an argument Narvinger has previously rejected on the grounds that Ericsson is not starting from scratch and facing substantial upfront costs, but instead building on years of activity. He and other senior executives at Ericsson also think CPUs and GPUs will prove costlier because of their need – as general-purpose chips – to support features that might never be used. Ericsson's latest earnings report, notably, came with yet another warning about rising component costs driven by the AI boom.

At the RAN-making mobile networks business group responsible for almost two thirds of Ericsson's revenues, the gross margin shrank by 0.2 percentage points year-over-year for the quarter, to 49.8%, as sales rose 9%, to 38.7 billion Swedish kronor (US$3.9 billion). But the unit's operating margin improved by 1.8 percentage points, to 21.8%, and it has been far more profitable than Nokia's equivalent business unit.

Yet to report third-quarter results, Nokia reported a second-quarter operating margin of 11.6% at its mobile infrastructure group, which includes core software and licensing activities besides the RAN business. That was down from 12.2% the year before. If Narvinger is right, a RAN strategy based on the use of GPUs may squeeze those margins in future quarters.

Nokia's deal with Nvidia might have been the fruit of economic necessity rather than about architectural roadmaps. Existing commercial partner Marvell is known to have had concerns about the returns it could generate from developing custom silicon products for Nokia after the Finnish company lost valuable US market share in a shrinking RAN sector. A deal with Intel seemed unlikely after Nokia's RAN relationship with the chipmaker was damaged by those initial 5G setbacks. In the absence of other viable suppliers, Nvidia might have been the only option Nokia had.