SoftBank unveils AI model speeding up 5G by 30%

SoftBank said it has developed a new AI architecture using Transformer models to improve wireless signal processing in 5G networks, achieving a 30% rise in uplink throughput in live tests.

The Tokyo-based telecom company reported that the technology was demonstrated in a real-time, over-the-air 5G environment, meeting industry latency standards and increasing throughput compared to its previous CNN-based model.

SoftBank said the new model processed signals with an average latency of 338 microseconds, about 26% faster than the earlier approach.

In simulation, the company said the architecture also improved downlink throughput by up to 31% for moving terminals, more than doubling rates seen with older models.

The Transformer-based system is designed to handle multiple signals processing tasks, potentially reducing development time and costs for future upgrades.

🔗 Source: SoftBank

🧠 Food for thought

1️⃣ Industry-wide momentum builds around AI-powered radio networks

SoftBank’s breakthrough comes amid significant industry collaboration on AI-RAN technology.

The AI-RAN Alliance, formed in February 2024 with founding members including AWS, Ericsson, Microsoft, and Nokia, has already surpassed 100 members in its first year 2.

This rapid growth reflects serious industry commitment, with over a third of telecom professionals now investing specifically in AI for RAN optimization according to NVIDIA’s recent survey 3.

The alliance’s three research focus areas—AI integration, operational efficiency, and new service development—align with what SoftBank has demonstrated: using advanced AI models to improve network performance and reduce operational complexity.

Major operators beyond SoftBank are also moving from concept to implementation, with Nokia announcing partnerships with KDDI and T-Mobile to advance AI-powered RAN technologies through real-world testing at Nokia’s AI-RAN Center in Dallas 4.

2️⃣ Technical breakthrough solves the performance-versus-speed trade-off

SoftBank’s achievement represents a significant technical milestone by simultaneously improving both AI performance and processing speed.

Their Transformer-based system achieved 338 microseconds average processing time—26% faster than their previous CNN approach—while also delivering superior performance gains 1.