- Japan’s SoftBank is quite proud of the AI RAN work it’s conducted thus far with Nvidia and others
- Part of that work is the AITRAS solution, which is supposed to bring cost savings and improved efficiency to the mobile network
- A SoftBank spokesman said its work on AITRAS is the real deal
MOBILE WORLD CONGRESS, BARCELONA – When SoftBank and Nvidia announced the formation of the AI-RAN Alliance a year ago, some folks questioned whether it was more or less a marketing gambit as opposed to a real engineering effort. After all, everybody was slapping “AI” on any and all marketing material to show how hip they are to the thing that’s sweeping the world.
Flash forward to this year, and SoftBank and friends are here to tell everyone who’ll listen that they’re serious as a heart attack when it comes to AI and the RAN, and it’s not just a marketing ploy.
“This is not a PowerPoint thing,” said Ryuji Wakikawa, head of the Research Institute of Advanced Technology at SoftBank, during a media briefing. “This is actual engineering work we've achieved.”
Last fall, SoftBank launched initiatives to develop AITRAS, a solution based on the AI-RAN concept that can host both AI and RAN workloads on the same Nvidia-accelerated computing platform. SoftBank not only plans to implement AITRAS across its own commercial network but wants to expand the solution to telecom operators globally starting in 2026.
This week, SoftBank announced, in collaboration with Fujitsu, the implementation of central unit (CU) functions on the Nvidia Grace CPU Superchip platform as part of efforts to accelerate the commercialization of AITRAS. In addition, they’re feting the implementation of distributed RAN (D-RAN) functionality to deploy both the CU and distributed unit (DU) functions on a single Nvidia GH200 Grace Hopper Superchip.
Of course, this would not be complete without the inclusion of Arm, the semiconductor company majority-owned by SoftBank that also attracted an investment from Nvidia. The Nvidia Grace CPU Superchip leverages two Arm Neoverse V2-based Grace CPUs, according to a SoftBank press release.
Through its R&D efforts, SoftBank said it achieved two types of architectures in AITRAS: centralized RAN (C-RAN), which deploys CU and DU on separate servers to enhance base station resource efficiency; and D-RAN, which processes CU and DU on the same server to minimize hardware configuration.
SoftBank expects to deploy C-RAN and D-RAN architectures in outdoor test scenarios between April 2025 and March 2026, with the next step being performance verifications.
Ericsson, Nokia, Red Hat collaborations
That’s not all. Separately, SoftBank is giving some love to its friends at Ericsson, Nokia and Red Hat.
SoftBank and Ericsson signed an MoU to expand their collaboration on advancing AI in the RAN, conducting joint research and development in their applied fields. Their work includes the demonstration of prototype Ericsson cloud RAN software running on the Grace CPU of Nvidia GH200 Grace Hopper Superchip.
With Nokia, SoftBank said it has integrated a new function into the AITRAS orchestrator. The new functionality allows AI and virtualized RAN (vRAN) to coexist on a single server with GPU while “dynamically optimizing resource allocation.” This implementation addresses the “AI and RAN” principle that is part of the overall AI-RAN concept.
As for Red Hat, SoftBank developed a solution that monitors energy usage and optimizes power consumption for vRAN and AI applications operating in AI-RAN data centers. Through this solution, AITRAS uses its orchestrator to dynamically allocate resources to AI applications based on power usage and other factors, all designed for achieving optimal power consumption.
AI for RAN
SoftBank also shared how it’s conducting research on “AI for RAN,” one of the concepts the AI-RAN Alliance is pursuing.
As part of its efforts to realize AI for RAN, SoftBank conducted verification tests on three use cases: “Uplink Channel Interpolation,” “Sounding Reference Signals Prediction” and “AI-driven MAC Scheduling.” The tests demonstrated the effectiveness of AI in improving RAN performance.
To put an even finer point on it, SoftBank said it conducted tests confirming that UL user throughput improved by about 20% in areas with poor network performance compared to conventional technology. Applying the AI technology to base stations, UL transmission speed degradation can be mitigated even in high-interference environments, making for a better user experience when it comes to uploading videos, images and other data.
SoftBank will be participating in demonstrations at several of its partner booths during MWC, including Arm, Fujitsu, Nokia and Red Hat, so keep an eye out for SoftBank’s AI-RAN on the show floor.
Stay up to date on all the news from Mobile World Congress with Fierce Network and FNTV.