It reported at home on July 20 during the 2025 China Unicom Partner Conference, China Unicom Research Institute, along with the Shanghai Artificial Intelligence Laboratory and other partners, who officially released the test results of July 19. With the rapid development of artificial intelligence technology and the speed of the season, the developmental model of the higher the highest-needed model, the development of the higher the highest-needed model, Development for high prevention, higher performance and high performance, high performance model,计算功率基础架构。但是，当前的计算能源显示出异构化和以上为中心的模式，并且面临诸如跨域资源连贯性，带宽和延迟瓶颈机制不完整的挑战，并延迟到较长的交付距离，以及有限的统一调度和Maexcellent功能。在这种背景下，中国Unicom以及上海人造情报实验室和行业合作伙伴，例如阿里巴巴云，启动，基本流动技术和贾吉亚技术，旨在GPU的国内异质环境，旨在培训电力整合需求的跨域碎片。 The hope of long-domains without loss of technology of delivery of the benefits of "computing power intelligent networking Ainet", accompanied by "Deeplink" super-large-scale cross-domain mixed with Shanghai training technology artificial intelligence, a cross-domain heterogenous mixed system of training with a length of more than 1,500 kiltero Shanghai Ling and Baoshan Doka in Baishan and Jinan, Shandong. Learned it inbahay mula sa kumperensya na sa pamamagitan ng awtomatikong pag-aayos ng pagsasaayos ng parameter ng modelo ng MOE at siksik na modelo sa ilalim ng ratio ng cross-domain na tagpo, ang pinagsamang koponan ay nakumpleto ang isang 100 bilyong parameter na malaking modelo ng异源Na Halo-Halong Pagsubok sa Pagsasanay ng higit sa 1500 kilmomtro sa buong域，在napatunayan na ang katumbas na kapangyarihan ng计算ay maaaring umabot umabot ng higit sa 95％ng katumbas na kapangyarihan ng组成一个单芯片和单个群集。在计算强度的层面上，中国Unicom考虑了内置的长距离RDMA协议堆栈的ROCE开关，并考虑到DC内和之间的高性能网络需求；在计算能力和资源管理的调度级别上，中国Unicom的“ Zhiyu”平台实现了统一的管理和调度，以及跨域网络和集群的工作以及PSCHEDUL的工作；在计算能力的协调水平上，基于基本的删除技术，例如加速训练，异质通信和并行方法，它实现了在超长距离异质GPU环境中对1000亿个参数大型模型的出色训练。中国Unicom说，上述成就标志着中国的巨大技术在主要领域的跨区域突破，例如调度计算能力基础架构，异源资源集成和高性能分布式培训的跨域协作，这为下一个互联网行业一代的发展奠定了基础。