VMO2 reports 20% GPU RAN gain as Ericsson cites 10%

Daniel Okafor

What did VMO2 measure in Paddington?

Virgin Media O2 recorded a 20% spectral-efficiency gain in a Nokia-site trial in Paddington, central London, using an Nvidia graphics processing unit within the radio access network (RAN). The site could carry roughly one-fifth more data per unit of spectrum without additional radio capacity or spectrum.

The operator installed the Nvidia processor at a busy Nokia site. The surrounding radio equipment comes entirely from Nokia and includes custom chips for conventional RAN algorithms, co-developed by Nokia and Marvell Technology.

The trial focused on the radio link between the mast antenna and customers' smartphones in Paddington.

Rob Joyce, VMO2's director of mobile access engineering, told reporters and analysts at a recent briefing: "What we've been working with them on is using AI to better estimate the channel quality."

Traditional algorithms rely on predefined radio parameters. Joyce said the AI model instead observes local conditions and tunes the basestation's algorithms to that environment in real time.

The trial followed a Nokia partnership with Nvidia struck about a year ago. Nvidia invested $1 billion in Nokia, while Nokia began redesigning its 5G products to run on GPUs instead of Marvell silicon. GPUs have become the default choice for developers training large language models in data centers.

Nokia says GPU-based RAN can improve spectral efficiency and deliver other operational benefits by replacing traditional algorithms with AI. The Paddington result supports the efficiency claim at one site, and VMO2 expects more substantial gains in future.

Could GPU RAN change operator spending?

Numerous other telcos are running GPU RAN trials this quarter, with commercial products due next year. Few have disclosed what they found. A result large enough to change network economics could persuade operators to fund GPU products instead of conventional upgrades.

VMO2 has already spent about £1.4 billion ($1.9 billion) on mobile network upgrades over the past two years, equivalent to roughly £700 million annually. It installed new radios from Nokia in southern England and Ericsson in the north.

A £700 million ($940 million) tranche announced early this year includes no GPU RAN deployment plans, Joyce confirmed. A mass GPU rollout across London and the rest of Nokia's footprint would still be a surprise.

The concern is power efficiency. Telcos and industry experts doubt that general-purpose GPUs can match purpose-built custom silicon. Any GPU RAN business must therefore produce benefits that outweigh extra power use and the cost of new hardware.

How does Ericsson's ASIC route compare?

Ericsson continues to market its application-specific integrated circuits (ASICs) as a suitable platform for AI. Around 60,000 basestations operated by its customers now support an AI successor to link adaptation, a RAN algorithm, in commercial networks.

CEO Per Narvinger said those deployments boost spectral efficiency by around 10%. That is Ericsson's vendor-reported result; VMO2's 20% figure came from a separate Nvidia GPU trial at one Paddington site.

Ericsson's high-capacity RAN Processor 6672 is designed for its own tower-side appliances, traditionally installed at the bottom of towers. Operators can also pool these ASICs in a centralized RAN facility supporting multiple radios.

Common off-the-shelf servers offer greater scope for aggregation. Dell, Quanta and Supermicro now incorporate RAN-capable GPUs into servers that could support cloud RAN deployments.

Joyce does not treat a GPU as a prerequisite for AI in the RAN. "You don't need a GPU to do this," he said, while adding that Ericsson's silicon is "very good as well."

What does centralization do to the fiber equation?

Centralization depends on fiber, not merely compute. Paddington has a 10 Gbit/s fiber connection, sufficient for local needs, but moving baseband processing to a central facility would require common public radio interface (CPRI) or evolved CPRI (eCPRI) links back to the radio sites. Each interface runs at 25 Gbit/s.

A design requiring three connections would need 75 Gbit/s to the tower, 7.5 times the site's present 10 Gbit/s link. VMO2 would have to add fiber or pay more for transmission.

Centralized GPUs could control several sites and coordinate scheduling across multiple technologies, which Joyce said could improve performance further. High-capacity fiber enables that architecture, but each additional connection adds cost and complexity.

Joyce's preferred route remains distributed RAN. Under that design, operators add baseband GPU resources at individual sites rather than build a central fiber pool. The Paddington result supports local GPU capacity, not yet a London-wide deployment mandate.

Where do the vendor roadmaps lead?

Nokia still sells products based on Marvell silicon and will do so for years. However, it no longer develops new technology with Marvell. A durable GPU RAN business therefore requires Nokia's customers eventually to accept new hardware.

Ericsson faces the reverse risk. Its current ASICs may not accommodate emerging AI features that today's more powerful GPUs can support. Operators benefiting from link adaptation on years-old equipment could face earlier replacement and lose some installed-base advantage.

That anxiety creates an opportunity for either Nordic supplier to gain market share. The constraint is vendor scarcity: with so few RAN vendors, operators need both Ericsson and Nokia to remain technically credible and commercially competitive.

The next test comes next year, when GPU RAN products are due to launch commercially. VMO2's 20% result gives operators evidence to evaluate, but its £700 million spending plan excludes GPU deployment, leaving power, fiber and hardware costs unresolved.