Softbank's Transformer AI model boosts 5G uplink throughput by 30%
Softbank developed a Transformer-based Artificial Intelligence (AI) model designed for wireless signal processing. This model enhanced uplink channel interpolation, enabling networks to assess the characteristics and current state of signal channels more effectively. Softbank reported an uplink throughput increase of approximately 20% using this new method compared to conventional signal processing methods. The demonstration was conducted on GPUs in a live Over-the-Air (OTA) wireless environment, confirming real-time operation alongside further throughput gains and ultra-low latency.
Technical Advancements:
The uplink channel interpolation with the new architecture improved throughput by approximately 8% against the traditional Convolutional Neural Network (CNN) model. In comparison to baseline methods without AI, the throughput saw an increase of around 30%, indicating ongoing enhancements in communication quality in real environments. Moreover, the model demonstrated an average processing time of about 338 microseconds, significantly under the required 1 millisecond for real-time 5G communications.
The Transformer model's success lies in its self-attention mechanism, allowing it to effectively grasp complex correlations in wireless signals over time. This capability improves performance without the need for normalization of input data, which preserves vital physical information that impacts communication quality. The versatile architecture can adapt to various tasks through minor adjustments, facilitating the development and deployment of AI solutions for different signal processing challenges.
Future Steps:
Softbank aims to commercialize these validated technologies, contributing to improvements in communication quality and network advancements using AI-RAN strategies. This development is poised to optimize the efficiency of capital expenditures for telecommunications, enhancing value through continuous software updates as advancements in AI models emerge.