SoftBank Corp. Demonstrates RAN Performance
Enhancement with AI Technology

Uplink Channel Interpolation

In the TDD (Time Division Duplex) method, which is primarily used in 5G, the number of available slots for Uplink (“UL”) transmission is fewer compared to Downlink (“DL”) transmission. This is particularly important as AI traffic generated by multi-modal generative AI applications is expected to have a higher UL to DL ratio compared to traffic generated by traditional voice and data applications. As a result, more efficient data transmission is required. To enhance performance, it is essential to improve the accuracy of channel estimation for signals received at base stations.

In mobile communication environments, radio conditions are fluctuating dynamically due to various factors such as time, location, network usage, and interference. As a result, improving channel estimation accuracy is becoming even more critical.

To enhance channel estimation accuracy using AI technology, SoftBank conducted a verification experiment in a lab environment using smartphones. A comparative test between conventional technology and AI technology confirmed that UL user throughput improved by approximately 20% in areas with poor network quality. With this technology applied to base stations, UL transmission speed degradation can be mitigated even in high-interference environments, enhancing the user experience when uploading videos, images, and other data.

This initiative was conducted in collaboration between by SoftBank, NVIDIA, and Fujitsu Limited (“Fujitsu”). NVIDIA provided its “ARC-OTA” (Aerial Research Collab - Over The Air) commercial testbed based on NVIDIA GH200 Grace Hopper Superchip for verification testing and customized the NVIDIA AI Aerial layer 1 software. Fujitsu and SoftBank designed and developed the AI, Fujitsu developed the embedded interface, and SoftBank coordinated the overall project, established the verification testing environment, and conducted the evaluations.

Additionally, we will showcase this demo at the AI-RAN Alliance section located at the Arm booth during MWC Barcelona 2025, starting from March 3rd.

Sounding Reference Signals Prediction

In 5G and 6G, advancements in base station technology are expected to enhance beamforming performance using Massive Antennas with an increased number of antenna elements. In TDD systems, beamforming generally relies on estimating the transmission path using Sounding Reference Signals (“SRS”) transmitted from user devices. SRS is sent from terminals to base stations at regular intervals; however, as the number of devices communicating with base stations increase, the SRS transmission interval becomes longer. Since transmission path estimation cannot be performed when SRS is not received, it is necessary to predict how SRS will change over time. However, when a device moves at high speed, channel fluctuations become more significant, making SRS prediction more challenging. If a prediction is inaccurate, beamforming based on incorrect information may result in degraded performance.

SoftBank conducted verification tests on communication performance improvement by utilizing AI technology to predict SRS during non-reception intervals. For the verification, an AI model based on the Multilayer Perceptron (“MLP”) algorithm was implemented in a system-level simulator. As a result, it was confirmed that the DL throughput of a device moving at 80 km/h improved by approximately 13%.

AI-Driven MAC Scheduling (MAC Scheduler)

With the enhanced beamforming performance enabled by massive antennas, cell capacity is expected to increase through user multiplexing with MU-MIMO (Multi-User Multiple Input, Multiple Output). However, to effectively perform multiplexing with MU-MIMO, multiple factors must be considered, including the correlation between users, wireless quality, and the prioritization of MAC-layer radio resource allocation in scheduling (MAC scheduling).

Additionally, the vast matrix computations required for different combinations of user devices are further increasing the complexity of processing.

SoftBank conducted a verification test to enhance performance through efficient user pairing by applying AI technology to MAC scheduling. Using the MLP (Multilayer Perceptron) algorithm for MAC scheduling in a system-level simulator, the results confirmed that the average user throughput improved by approximately 8%.

Furthermore, SoftBank developed a system-level simulator as a verification platform to efficiently evaluate the enhancement of RAN performance through AI technology. By utilizing this simulator, various AI-driven effects can be tested efficiently, enabling more effective verification of AI applications in RAN.
Going forward, SoftBank will continue to advance the development of AI-driven communication technologies, striving to deliver even higher-quality communication services.

Ryuji Wakikawa, Vice President, Head of the Research Institute of Advanced Technology at SoftBank said: “‘AI for RAN’ which SoftBank is promoting, demonstrates the significant impact AI can have on enhancing RAN performance. Being able to achieve such substantial performance improvements in implementation without requiring changes to communication specifications for AI utilization suggests the potential for major evolution of our infrastructure through AI innovation in the telecommunications industry. SoftBank will continue to drive innovation, with the goal of providing our customers with the best communication experience.”

Soma Velayutham, Vice President of Telecoms, NVIDIA, said: “AI embedded into radio signal processing with a software defined platform will deliver transformative gains and set continuous performance and efficiency benchmarks not achievable with traditional techniques. SoftBank's innovations in AI for RAN, harnessing NVIDIA AI Aerial, mark a significant milestone in progressing AI-RAN technology and underscore its ability to deliver exceptional performance and efficiency with AI”

Shingo Mizuno, Corporate Executive Officer & EVP, Vice Head of System Platform (in charge of Network Business), Fujitsu Limited, said: “Fujitsu has been engaged in the development of high-performance vRAN software utilizing GPUs. In our recent joint research with SoftBank and NVIDIA, Fujitsu's AI technology contributed to the improvement of uplink performance, which we believe is a significant achievement towards the enhancement of wireless networks through the convergence of future wireless and AI technologies. We hope that the development of new technologies, including AI, will continue to provide value to even more customers and society as a whole.”