SoftBank and Nvidia have piloted the world’s first artificial intelligence (AI) and 5G telecoms network in Japan.
The network can run AI and 5G workloads concurrently, a process known as an artificial intelligence radio access network (AI-RAN).
SoftBank, as part of its strategy to boost customer experience, has demonstrated that its NVIDIA-accelerated AI-RAN solution has achieved carrier-grade 5G performance and was able to do so while using the network’s excess capacity to run AI inference workloads concurrently.
SoftBank has used NVIDIA AI Enterprise to build AI inference applications, including autonomous vehicle remote support, robotics control and multimodal retrieval-automated generation at the edge. All inference workloads were able to run optimally on SoftBank’s AI-RAN network.
SoftBank’s software-defined 5G radio stack is optimized for NVIDIA’s AI computing platform and includes L1 software enhanced by SoftBank based on NVIDIA Aerial CUDA-accelerated RAN libraries.
SoftBank will incorporate NVIDIA Aerial RAN Computer-1 systems, which it estimates can use 40 percent less power than traditional 5G network infrastructure, into its solution moving forward.
NVIDIA and SoftBank partners that contributed to the trial of SoftBank’s AI-RAN solution include Fujitsu and Red Hat.
SoftBank aims to build an ecosystem that connects the demand and supply of AI technology by using NVIDIA AI Enterprise serverless application programming interfaces and its in-house developed orchestrator. This enables SoftBank to dispatch external AI inferencing jobs to an AI-RAN server when computing resources are available to deliver localized, low-latency, secure inferencing services.
“Shifting from single-purpose to multi-purpose AI-RAN networks can mean 5x the revenue for every dollar of capex invested,” said Ronnie Vasishta, senior vice president of telecom at NVIDIA. “SoftBank’s live field trial marks a huge step toward AI-RAN commercialization with the validation of technology feasibility, performance and economics.”
Nvidia CEO Jensen Huang said SoftBank was the first to receive its new Blackwell chip designs, which it is incorporating into the AI supercomputer it is developing to create its own generative AI model.
Junichi Miyakawa, CEO of SoftBank, said: “With our extremely powerful AI infrastructure and our new, distributed AI-RAN solution ‘AITRAS’ that reinvents 5G networks for AI, we will accelerate innovation across the country and throughout the world.”
SoftBank is slated to receive the world’s first NVIDIA DGX B200 systems, which will serve as the building blocks for its new NVIDIA DGX SuperPOD supercomputer.
SoftBank plans to use its Blackwell-powered DGX SuperPOD for its own generative AI development and AI-related business, as well as that of universities, research institutions and businesses throughout Japan.
SoftBank’s DGX SuperPOD is expected to be Japan’s most performant to date. Featuring NVIDIA AI Enterprise software and NVIDIA Quantum-2 InfiniBand networking, it is also ideal for the development of large language models.
In addition to its DGX SuperPOD, SoftBank plans to build another NVIDIA-accelerated supercomputer to run compute-intensive workloads. Initial plans for the supercomputer are based on an NVIDIA Grace Blackwell platform design featuring NVIDIA GB200 NVL72 multi-node, liquid-cooled, rack-scale systems that combine NVIDIA Blackwell GPUs with power-efficient Arm-based NVIDIA Grace CPUs.
NVIDIA and SoftBank estimate that telco operators can earn roughly $5 in AI inference revenue from every $1 of capex it invests in new AI-RAN infrastructure. Taking into account its opex and capex costs, SoftBank estimates it can achieve a return of up to 219 percent for every AI-RAN server it adds to its infrastructure.
