Vendors & EquipmentAI RanRan SiliconGpusVodafone
Vodafone Tests AI-RAN Chips in Malaga but Holds Off on GPUs
Vodafone is testing AI-RAN on multiple chip types in Malaga and sees major gains, but has given no green light to GPU deployments as cost and hardware concerns persist.
Vendors & EquipmentWhy it matters
- Vodafone has no green light for GPU deployments, testing AI-RAN chips at its Malaga campus
- The global RAN products market was worth as little as $35 billion last year, per Omdia
- Nokia took a $1 billion Nvidia investment a year ago and will base future RAN products on its GPUs
- A Nokia–Virgin Media O2 AI-RAN demo on Marvell silicon showed 20% spectral efficiency gains
- Nokia has activated only 10 of 7,000 VodafoneThree sites awarded in September last year
The story
Vodafone is testing AI-RAN software on different chip architectures at its Malaga innovation campus but has not approved GPUs for network deployment, according to Kyriakos Exadaktylos, the operator's head of RAN evolution.
Speaking at a recent AI-RAN symposium hosted by Omdia, a Light Reading sister company, Exadaktylos backed the performance case for AI in the radio access network while withholding judgment on the hardware. "I think the route to improve is quite significant, and AI will help for that," he said.
The work is under way in the Spanish port city, where Vodafone has replaced the historical algorithms used in the RAN with more adaptive AI software and then run that software on various chips. The campus sits a kilometer from Malaga's old town. Findings could eventually shape the type of network Vodafone builds in the UK.
What does this change for silicon choice?
Mobile networks have run on custom silicon purpose-built for RAN signal processing since the mass-market era. Alternatives emerged in the late 4G period. Intel has pitched general-purpose CPUs as a substitute. Arm-based CPUs have been proposed more recently. The newest idea is an Nvidia GPU — the chip most associated with AI workloads.
Vodafone is not yet convinced. "We need to go into our labs that we're now building with advanced capabilities to test the benefit on network efficiency," said Exadaktylos in response to questions from Light Reading. "So, we have not, at the moment, stamped that we have a green light to go to deployments with GPUs."
Operator resistance carries two main drivers:
- GPUs carry a reputation for high power consumption and cost, despite Nvidia's efforts to market lighter RAN-suitable designs. Several telco executives doubt they would be economical.
- GPUs would force new hardware investment, even if limited to slotting a card into an existing appliance. No commercially sensible operator buys new silicon if installed chips deliver equal benefits.
There is also a technical objection. RAN processing must execute quickly, so its AI models must stay comparatively small. Experts are skeptical a GPU would outperform custom silicon designed for whatever AI needs arise. Google, for its own data center AI, has been shifting from GPUs to tensor processing units, a type of ASIC.
What is the case for custom silicon?
Ericsson, one of Vodafone's preferred vendors, runs an AI-native form of link adaptation on internally designed ASICs. "The really critical part here is that you have about 40 microseconds before you need to make a decision on link optimization," said Gabriel Foglander, Ericsson's head of RAN strategy. "The inferencing speed is really critical — and the deterministic nature of that."
Yet custom RAN chips serve a small market. Omdia research puts the entire RAN products market at as little as $35 billion worldwide last year, leaving ROI for purpose-built silicon looking poor against general-purpose chips.
Rapid AI-RAN software change could force more frequent ASIC refreshes. A powerful GPU, though over-dimensioned today, might survive longer and reduce future hardware spending. "What is going to determine that is the ROI that you're going to see from a kind of faster cadence of replacing baseband," said Foglander, noting that truck rolls and commissioning form a substantial part of capex.
When might GPUs be needed?
Exadaktylos sees software headroom in today's hardware. Existing algorithms with AI embedded "will not need a change on silicon," he said. More advanced "superintelligent" algorithms could have an impact, but "this is not today. Maybe it's two, three years down the road, and we will need to validate."
Telco hesitancy would hinder Nvidia's RAN progress at an awkward moment, following the abrupt exit of CEO Jensen Huang last week. The sharper pain could fall on Nokia, which took a $1 billion Nvidia investment a year ago and plans to base future RAN products exclusively on its GPUs.
Nokia is expected to keep supporting its latest Marvell Technology co-developed products in commercial networks for several years. A recent AI-RAN demo with the UK's Virgin Media O2 on that Marvell silicon showed a 20% improvement in spectral efficiency — the same figure Nokia had previously cited for GPUs this year. Nokia has brandished larger numbers it believes achievable with Nvidia by 2028, though the baseline is unclear.
Exadaktylos notes that if gains were not plausible through OFDM, the waveform at the heart of 5G, the industry would not have kept it for 6G. "If we had reached the limit of the modulation scheme, then 6G would have completely changed that," he said.
Outside the UK, Vodafone relies mainly on Huawei and Ericsson in Europe, with Nokia historically limited to Albania. Nokia landed a 7,000-site deal with VodafoneThree in September last year and had activated only ten sites by April. What Vodafone learns in Malaga over the coming years could determine the silicon at the heart of that UK build.
Also reported
Original: theguardian.com
More from James Calloway
Show full bio
Staff writer covering companies, deals and strategy in Carrier business, 5G and 6G, fiber and broadband at Telecom Gazette.
524 articles