China Mobile Reveals Innovative Intelligent Computing Interconnection Technology and Unveils the World's Pioneering 100T-Level Intelligent Computing Interconnection Device Prototype
15 hour ago / Read about 0 minute
Author:小编   

On March 3 (local time on March 2), at the Mobile World Congress (MWC) in Barcelona, Spain, China Mobile proudly presented its groundbreaking Scale-Across technology—GSE-DCI (Global Scheduled Ethernet Intelligent Computing Data Center Interconnection)—and showcased the world's inaugural ultra-100T intelligent computing interconnection router prototype, boasting an impressive throughput of 115.2Tbps. This milestone launch signifies a major leap forward in China's core interconnection network technology for intelligent computing centers.

At present, a single intelligent computing center finds it challenging to keep pace with the soaring computing demands, primarily due to limitations in power and space. Consequently, constructing super clusters by interconnecting multiple intelligent computing centers has emerged as a pivotal technological approach. China Mobile's GSE-DCI technology is specifically designed to tackle the hurdles encountered in intelligent computing interconnection links, including multi-wavelength load balancing, long-distance congestion management, security safeguards, and coordinated computing network scheduling.

The router prototype unveiled by China Mobile boasts exceptional features such as high density, long-distance transmission capabilities, efficiency, ultra-broad bandwidth, lossless performance, and robust security. It significantly elevates the computing efficiency of distributed AI training across vast distances of hundreds of kilometers, achieving over 98% of the efficiency of a single-node cluster. This makes it well-suited to meet the training and inference requirements for models with tens of trillions of parameters. Ultimately, this technological advancement provides robust support for intelligent computing scenarios, including cross-regional collaborative training and compute-storage separation.