Japan’s SoftBank is the marquee customer of Nvidia’s latest generative artificial intelligence (AI) work targeting 5G and 6G applications, with that 5G support coming “essentially … for free.”

The work will involve SoftBank using Nvidia’s Grace Hopper chip to power its data centers. These will house the telecom giant’s AI and wireless applications – including support for its 5G and virtualized radio access network (vRAN) – using a multi-tenant common server platform designed to reduce costs and lower power consumption.

Grace Hopper is Nvidia’s latest chip architecture and is targeted at AI and high-performance computing (HPC) workloads. It was initially unveiled in 2021 and taps into Arm’s Neoverse V2 core design.

SoftBank is using that platform alongside Nvidia’s BlueField-3 data processing units (DPUs) to boost its software-defined 5G vRAN and generative AI applications. The carrier is also using Nvidia’s Spectrum Ethernet switch to provide 5G timing protocol.

These run on an Nvidia-accelerated 1U MGX-based server design that can support downlink throughput speeds up to 36 Gb/s.

Ronnie Vasishta, SVP for telecom at Nvidia, explained during a press briefing that the platform’s software-defined nature allows customers to manage their server usage.

“It can take full advantage of the cloud economics of a data center. It can run in an AI factory, whether it be a public cloud,  a distributed cloud, an on-prem cloud – it's the same hardware. It's a software update if you want to reconfigure that system,” Vasishta said. “It's multi-tenant, meaning that cloud can run AI, it can run RAN as a tenant in the cloud. And many of these clouds … are getting built because of the dramatic growth of AI needs and AI applications. And so now you're essentially getting 5G for free.”

Nvidia claims this programmability allows users to gain up to a four-times improvement in return on investment when running vRAN workloads compared with a single-purpose 5G vRAN. Many of those single-purpose 5G vRAN deployments run on ubiquitous x86 architectures.

“We're now able to run 5G as pure software-defined,” Vasishta added. “Think of it as RAN in the cloud or RAN in an AI factory. You get best-in-class performance. You get open and flexible. The RAN stack can be updated, can be reconfigured purely through software.”

Despite that performance potential, Vasishta did note that the platform might be overkill for operators that are just looking to support vRAN deployments. “There’s cheaper ways to go implement that,” he said.

“We're talking to many telcos today that have sites where they have power, they have racks in those sites and they want to leverage those sites more cost efficiently for other applications as well, or more pooling if RAN,” Vasishta said. “I think that model is becoming more prevalent to run, more computational pooling of RAN sites and RAN traffic rather than just the single DU.”

Nvidia had previously helped SoftBank construct an AI-to-5G Lab that uses Nvidia’s hardware, and its vRAN and AI processing middleware; virtualized radio signal processing software from Mavenir; and core network software, AI image processing application software on mobile edge compute (MEC), and physical radio units (RUs) provided by Foxconn Technology.

SoftBank earlier this year used that lab to stream an image using a 5G signal running over vRAN components powered by GPUs between two devices that used a MEC AI application to detect the person in the image.

CORRECTION: This story has been updated to correct that the Grace Hopper chip architecture taps into Arm’s Neoverse V2 core design.