AI智算发展对高速光模块的应用需求研究  

Study on Application Demand of High-speed Optical Modules for AI Intelligent Computing Development

作　　者：栾昊立 王晓东[1] 杨锐 郝建宇 赵铭浩 尹祖新 王丽琼[1] Luan Haoli;Wang Xiaodong;Yang Rui;Hao Jianyu;Zhao Minghao;Yin Zuxin;Wang Liqiong(China Unicom Research Institute,Beijing 100048,China;China United Network Communications Group Co.,Ltd.,Beijing 100033,China)

机构地区：[1]中国联通研究院,北京100048 [2]中国联合网络通信集团有限公司,北京100033

出　　处：《邮电设计技术》2024年第6期7-11,共5页Designing Techniques of Posts and Telecommunications

摘　　要：AI智算技术的高速发展驱动高速光模块需求量激增,大规模数据处理、大模型训练和推理等任务对高速光模块提出了前所未有的高要求。通过分析大模型训练的分布式并行计算需求,建立通信模型,并以GPT-3为例定量分析大模型通信量,由于通信量巨大,完成大模型训练的数据通信时间远高于并行计算时间。因此,在不降低计算性能的前提下,降低通信时间成为AI智算对通信网络的核心诉求,而采用更高速率的光模块互联、提升有效带宽是解决问题的主要途径。AI智算对高速光模块技术的需求将主要体现在更高速率、更大规模、高集约化、低功耗、高稳定性以及可管可控等方面。The rapid development of AI smart computing technology drives the surge in demand for high-speed optical modules,and tasks such as large-scale data processing,large model training and reasoning put forward unprecedentedly high requirements for high-speed optical modules.A communication model is established by analyzing the distributed parallel computing requirements for large model training,and quantitatively analyze the large model communication volume by taking GPT-3 as an example.Due to the huge communication volume,the data communication time for completing large model training is much higher than the parallel computing time.Therefore,under the premise of not losing computing performance,reducing communication time becomes the core demand of AI computing on communication network,and adopting higher speed optical module interconnection and improving effective bandwidth is the main way to solve the problem.The demand of AI computing on high-speed optical module technology will be mainly embodied in the aspects of higher speed,larger scale,high intensification,low power consumption,high stability,and manageable and controllable.

关 键 词：AI 智算 大模型 GPT-3 高速光模块 

分 类 号：TN913.7[电子电信—通信与信息系统]