Sugon has formally rolled out the scaleFabric Token Acceleration Technology System, pioneering the large-scale deployment of IBGDA technology in China. This innovative system boosts token throughput by establishing a tightly integrated, three-tier acceleration channel encompassing 'computing, storage, and transmission', thereby facilitating a high-speed, end-to-end data flow. It extends path optimization to large-scale clusters, significantly enhancing token generation efficiency through collaborative efforts.

Against the backdrop of the MoE (Mixture of Experts) architecture becoming the de facto standard for large-scale model inference, the IBGDA technology minimizes CPU intervention in critical communication pathways, thereby elevating parallel communication efficiency. Historically, large-scale implementations of this technology were predominantly led by international vendors, with limited domestic adoption.

However, Sugon's self-developed IBGDA technology has now successfully completed verification in a massive ten-thousand-card cluster, effectively bridging the domestic gap in this area. As the technology progresses towards widespread deployment, the associated Token Acceleration Technology System is anticipated to enhance data exchange efficiency across multiple cards. This breakthrough offers a domestic high-speed interconnection solution tailored for MoE models and other applications, ultimately propelling the performance of China's intelligent computing infrastructure to new heights.