SoftBank Corp.'s AI-RAN Realizes
"Software-only Massive MIMO" on GPU

Outline of the Outdoor Field Trial

In this outdoor field trial, AITRAS utilized the GPU's highly efficient parallel processing capability to execute large-scale matrix computations required for Massive MIMO and PHY-layer signal processing fully in software on GPU within the Distributed Unit (DU).
As a result, SoftBank confirmed stable operation of 16-layer MU-MIMO DL in outdoor conditions.

Compared with the conventional 4-layer configuration, both spectral efficiency and throughput improved approximately threefold. Even under high-traffic conditions, per-user throughput remained stable with high-quality communication, leading to greater overall base-station capacity.

This achievement marks a large-scale validation in which PHY-layer processing was executed fully in software on GPU, enabling stable Massive MIMO operation within the RAN's specified processing time. It represents an important technical milestone toward the commercialization of AI-RAN.

System Architecture and Features of AITRAS

In AITRAS, the interface between the Distributed Unit (DU) and the O-RAN(Open Radio Access Network)-compliant Radio Unit (O-RU) conforms to O-RAN Split Option 7.2x. While some Massive MIMO implementations offload all or parts of the uplink channel estimation and equalization processing from the DU to the O-RU, AITRAS adopts a fully software-based approach within the DU by leveraging GPU computing powers. This architecture enables the realization of a Massive MIMO ecosystem by utilizing general-purpose O-RUs without adding any new functions to them.

By aggregating uplink radio signals from multiple O-RUs into a single DU equipped with an integrated AI processing unit, AI-based coordinated control for radio signal quality can be effectively implemented across multiple O-RUs. This allows for enhanced radio signal quality through Uplink Channel Interpolation^* and increased capacity via advanced beamforming techniques.

Based on 3GPP and O-RAN standards, this advanced and fully software-based RAN architecture is designed to maximize RAN capabilities using AI.

[Note]

1. *
   Uplink Channel Interpolation is a signal processing technology that estimates and interpolates the state of the transmission path (channel) over which data is sent, based on the Demodulation Reference Signal (DMRS), which is allocated to only a part of the communication resources. It achieves faster, more stable communication by accurately determining the characteristics of the entire transmission path from this limited measurement information.

Future Outlook

Through the development of AITRAS, SoftBank will further accelerate its efforts toward practical implementation of AI-RAN, driving the creation of sustainable and flexible next-generation network infrastructure. SoftBank will continue conducting field trials and aims to introduce AITRAS into its commercial network from 2026 onward.

This initiative is being demonstrated at SoftBank's booth during NVIDIA GTC Washington, D.C., which begins on October 27, 2025.

Ryuji Wakikawa, Vice President and Head of the Research Institute of Advanced Technology at SoftBank Corp., said: "This field trial is an important achievement demonstrating that GPU-based AI-RAN can fully execute the radio signal processing required for Massive MIMO—demanding both high bandwidth and ultra-low latency—entirely in software. By combining AI and GPU computing, we will continue to advance development toward realizing SoftBank's AI-RAN vision, which aims to build more flexible and sustainable next-generation networks."

Soma Velayutham, Vice President, Telecoms at NVIDIA, said:"A Software-defined and AI-native architecture is essential for future wireless networks to enable rapid innovation and highest spectral efficiency. Gains with AI will be even greater for higher-order MIMO processing, such as ultra MIMO. SoftBank's performance gains achieved using a fully software-defined massive MIMO implementation of AI-RAN in collaboration with NVIDIA validates how accelerated computing is revolutionizing radio access networks."