Ericsson bakes neural accelerators into radios

Ericsson has launched a suite of AI-ready radios, antennas, and RAN software features designed to meet what it describes as AI-driven shifts in network demand, with a particular focus on uplink performance and latency.

At the hardware layer, Ericsson is integrating neural network accelerators into Ericsson Silicon for Massive MIMO radios. The company describes these accelerators as programmable matrix cores integrated within its Many-Core Architecture, intended to boost on-site AI inference inside radio units and enable distributed optimisation across the stack.

The product rollout is framed around three pillars: ten AI-ready radios, RAN software enhancements, and five high-performing antennas. On the radio side, Ericsson is highlighting new Massive MIMO and remote radios engineered for higher downlink efficiency and uplink gains. Named products include AIR 3286 for high-power FDD Massive MIMO, and AIR 3211, which combines Massive MIMO on both TDD and FDD. Ericsson is also expanding an eight-receiver portfolio via Radio 4891 and Radio 4458, positioned to lift uplink performance for AI and AR applications.

Further hardware additions include high-power triple-band radios, Radio 4488 and Radio 4464, aimed at network consolidation and RAN sharing, alongside TDD Massive MIMO updates such as AIR 3267, specified with 600 MHz instantaneous bandwidth in a 13 kg unit, and AIR 6492 with 480 W output power and 256 antenna elements.

On the software side, Ericsson is adding AI-managed Beamforming, AI-powered Outdoor Positioning, and what it calls a state-of-the-art AI model for instant coverage prediction, complementing its AI-native Link Adaptation software. To meet bounded-latency requirements, Ericsson is also introducing a Latency Prioritized Scheduler and Low Latency Mobility, which it says can deliver up to seven times faster response times.

The antenna updates cover interleaved and passive antenna designs intended to improve spectrum utilisation and simplify site design. Ericsson says the new energy-efficient passive antennas are built on a “trio net” design to maximise uplink performance, carrier aggregation, and spectral efficiency. It is also expanding its Interleaved AIR portfolio with additional configurations for TDD and FDD Massive MIMO deployments.

Mårten Lerner, Head of Networks Strategy and Product Management at Ericsson, said: “As AI transforms traffic patterns and raises consumer expectations, networks must provide precise performance where and when it’s needed most. At Ericsson, we’re committed to delivering that needed performance. We are also embedding AI solutions across our full portfolio, introducing AI RAN software solutions that deliver revolutionary improvements in spectral efficiency. We are now taking the final step to full AI enablement across our portfolio by introducing neural network accelerators in our leading Massive MIMO portfolio.”

Ericsson also pointed to operator deployment activity, with Iain Milligan, Network Development & Infrastructure Director at VodafoneThree, stating: “By bringing the Vodafone and Three networks together, on a single, high-performing 5G infrastructure powered by Ericsson’s RAN, VodafoneThree is ushering in a new era of connectivity — unlocking next generation technology and paving the way for AI-driven innovation. We’re embedding AI throughout our network, reducing operational complexity, boosting energy efficiency, and delivering faster, more reliable connectivity for our customers in every corner of the country.”

The immediate engineering challenge is less about adding AI labels, and more about whether embedded inference, scheduling changes, and new radio variants can deliver measurable uplink and latency gains without inflating power budgets or complicating RF site integration. Ericsson’s approach is to push acceleration and optimisation closer to the radio edge, then pair it with software features that translate into predictable service behaviour.