In China, a novel distributed intelligent network exceeding 2000 km in length has been launched. It integrates disparate computing centers into a unified structure, exhibiting an impressive 98% performance level, reports Science and Technology Daily.
According to the project head, academician Liu Yunjie, this infrastructure is intended to expedite the training of large-scale artificial intelligence models, as well as support telemedicine and industrial internet applications. He stressed the network’s capability to swiftly handle tasks where minimizing latency time is critically important.
Training a model possessing hundreds of billions of parameters and requiring over half a million iterations takes roughly 16 seconds per cycle using the new network. Previously, a similar operation required more than 20 seconds. During a trial run, the system transmitted 72 terabytes of data from the massive FAST radio telescope over a thousand kilometers in just 1.6 hours, whereas transferring the same volume of data via standard internet connection would have taken almost two years.
This platform forms part of the National Experimental Infrastructure of Future Networks, deployment of which commenced in 2013. Currently, it spans 40 population centers and incorporates over 55 thousand kilometers of fiber-optic lines.
About the Author
