|
On 21 August 2025, SoftBank announced that, by leveraging the power of transformer AI architecture for wireless signal processing, it achieved a 30% increase in 5G throughput in live demonstrations compared to baseline non-AI methods. In contrast, traditional convolutional neural networks (CNNs) improved uplink throughput by only roughly 20%. The demonstration, conducted in a 3rd Generation Partnership Project (3GPP)-compliant live wireless environment, proved real-time performance and ultralow latency, validating that transformers can effectively optimize communication quality in real-world 5G networks.
The use of transformers allows for capturing wide-ranging signal correlations across frequency and time, unlike CNNs, which focus on local input areas. Transformers also allow handling complex patterns caused by interference and reflection, improving accuracy. The model processes raw amplitude signals, retaining vital physical data for enhanced channel estimation, and can be easily adapted for different RAN tasks such as channel interpolation, sounding reference signal prediction, and demodulation with minor output adjustments.
Full Article: IEEE Vehicular Technology Magazine, Volume 21, Number 1, April 2026
|
