SoftBank Corp. Boosts 5G AI-RAN Throughput By 30% With New Transformer AI

Achieves groundbreaking low-latency architecture for signal processing

SoftBank Corp. ("SoftBank") announced the development of a new AI architecture that leverages the high-performance "Transformer" AI model for wireless signal processing as part of its research and development of "AI for RAN," a key concept of AI-RAN that utilizes AI to advance Radio Access Networks (RANs). SoftBank achieved a groundbreaking technological advancement that delivers an approximately 30% improvement in 5G throughput.

SoftBank successfully demonstrated both real-time operation in a live wireless environment compliant with 3GPP 5G standards and a significant improvement in communication quality. This achievement marks a major step forward for AI-RAN, moving it from the conceptual phase to practical implementation.

Demonstration in a real-world environment: uplink throughput jumps approx. 30% with ultra-low latency
SoftBank is advancing its R&D for RAN enhancement through a phased approach. In previous research using a Convolutional Neural Network (CNN), a type of AI model, for "uplink channel interpolation,"^*1 SoftBank successfully increased uplink throughput by approximately 20% compared to a conventional signal processing method (the baseline method).^*2 In the latest demonstration, the new Transformer-based architecture was run on GPUs and tested in a live Over-the-Air (OTA) environment. In addition to confirming real-time operation, the results showed further throughput gains and achieved ultra-low latency.

• Significant throughput improvement:
  Uplink channel interpolation using the new architecture improved uplink throughput by approximately 8% compared to the conventional CNN model. Compared to the baseline method without AI, this represents an approximately 30% increase in throughput, proving that the continuous evolution of AI models leads to enhanced communication quality in real-world environments.

• Higher AI performance with ultra-low latency:
  While real-time 5G communication requires processing in under 1 millisecond, this demonstration with the Transformer achieved an average processing time of approximately 338 microseconds, an ultra-low latency that is about 26% faster than the CNN-based approach. Generally, AI model processing speeds decrease as performance increases. This achievement overcomes the technically difficult challenge of simultaneously achieving higher AI performance and lower latency.

Demonstration in a simulation environment: downlink throughput improvement rate more than doubled
Using the new architecture, SoftBank conducted a simulation of "Sounding Reference Signal (SRS) prediction,"^*3 a process required for base stations to assign optimal radio waves (beams) to terminals. Previous research using a simpler Multilayer Perceptron (MLP) AI model for SRS prediction confirmed a maximum downlink throughput improvement of about 13% for a terminal moving at 80 km/h.^*2

In the new simulation with the Transformer-based architecture, the downlink throughput for a terminal moving at 80 km/h improved by up to approximately 29%, and by up to approximately 31% for a terminal moving at 40 km/h. This confirms that enhancing the AI model more than doubled the throughput improvement rate. This is a crucial achievement that will lead to a dramatic improvement in communication speeds, directly impacting the user experience.

Technical challenges and features of the new architecture
The most significant technical challenge for the practical application of "AI for RAN" is to further improve communication quality using high-performance AI models while operating under the real-time processing constraint of less than one millisecond. SoftBank addressed this by developing a lightweight and highly efficient Transformer-based architecture that focuses only on essential processes, achieving both low latency and maximum AI performance. Its key features are as follows:

1. Grasps overall wireless signal correlations
   By leveraging the "Self-Attention" mechanism, a key feature of Transformers, the architecture can grasp wide-ranging correlations in wireless signals across frequency and time (e.g., complex signal patterns caused by radio wave reflection and interference). This allows it to maintain high AI performance while remaining lightweight. Convolution focuses on a part of the input, while Self-Attention captures the relationships of the entire input.
2. Preserves physical information of wireless signals
   While it is common to normalize input data to stabilize learning in AI models, the architecture features a proprietary design that uses the raw amplitude of wireless signals without normalization. This ensures that crucial physical information indicating communication quality is not lost, significantly improving the performance of tasks like channel estimation.
3. Versatility for various tasks
   The architecture has a versatile, unified design. By making only minor changes to its output layer, it can be adapted to handle a variety of different tasks, including channel interpolation/estimation, SRS prediction, and signal demodulation. This reduces the time and cost associated with developing separate AI models for each task.

The demonstration results show that high-performance AI models like Transformer and the GPUs that run them are indispensable for achieving the high communication performance required in the 5G-Advanced and 6G eras. Furthermore, an AI-RAN that controls the RAN on GPUs allows for continuous performance upgrades through software updates as more advanced AI models emerge, even after the hardware has been deployed. This will enable telecommunication carriers to improve the efficiency of their capital expenditures and maximize value.

Moving forward, SoftBank will accelerate the commercialization of the technologies validated in this demonstration. By further improving communication quality and advancing networks with AI-RAN, SoftBank will contribute to innovation in future communication infrastructure.

^*1 Uplink channel interpolation: A signal processing technique that estimates and interpolates the state of the transmission path (channel) for data transmission areas from reference signals (DMRS) placed only in certain parts of the communication resource. It contributes to faster and more stable communication by accurately understanding the characteristics of the entire transmission path from limited measurement information.

^*2 For more information, please see this press release dated March 3, 2025: "SoftBank Demonstrates RAN Performance Improvement with AI Technology."

^*3 Sounding Reference Signal (SRS) prediction: A technique that prevents degradation in beamforming performance by estimating the transmission path from previously received SRS data at times when the base station cannot receive the SRS sent from the terminal at regular intervals