一种基于矩阵运算的智算中心参数网络部署方法  

A method for deploying parameter network of artificial intelligent computing centers based on matrix operations

作　　者：王祎玮 田雯 刘鹏 董江帆 魏逸哲 WANG Yi-wei;TIAN Wen;LIU Peng;DONG Jiang-fan;WEI Yi-zhe(China Mobile Group Design Institute Co.,Ltd.,Beijing 100080,China)

机构地区：[1]中国移动通信集团设计院有限公司,北京100080

出　　处：《电信工程技术与标准化》2024年第S02期23-28,共6页Telecom Engineering Technics and Standardization

摘　　要：随着生成式人工智能(AIGC)技术的发展,各类大模型层出不穷,而智算资源池网络是进行模型训练的关键环境。如何快速搭建、部署一个规模化的智算中心网络是能否快速形成大模型应用能力的关键一步。本文提出了一种基于矩阵运算的智算中心参数网络部署方法,通过矩阵运算的方法实现智算资源池中参数平面的交换机之间、交换机和服务器之间的快速匹配,将组网模式标准化,极大地提升了部署参数网络的效率。With the development of artificial intelligence generated content,all kinds of large models are emerging,and the network of the intelligent computing resource pool is the key environment for model training,how to quickly build and deploy a scaled artificial intelligent computing center network is the key step to form a large model quickly.In this paper,we propose a deployment method of ICC parameter network based on matrix operation,which realizes rapid matching between RDMA switches,and between RDMA switches and servers in the parameter plane in the artificial intelligent computing center resource pool,standardizes the networking mode,and greatly improves the eficiency of deploying parameter network.

关 键 词：参数网络 RDMA交换机 交换机部署 智算中心 

分 类 号：TN915[电子电信—通信与信息系统]