AI大模型场景下智能计算技术选型分析

Selection and analysis of intelligent computing technology in AI large-scale model scenarios

随着ChatGPT的诞生,AI大模型相关的产品和服务呈爆发式增长,国内互联网企业和运营商等纷纷加入“百模大战”。现有云平台中智能服务器采用传统GPU配置规格,无法满足大模型训练所需要的高性能算力和卡间互联带宽需求。本文主要研究扣卡模组形态的智能计算芯片技术,以及整机服务器的散热技术等内容,为后续建设智算中心提供选型依据。

With created of ChatGPT,products and services related to AI large-scale models have exploded.Domestic Internet companies,telecom operators,and others have joined the“hundred-model war”.The intelligent servers in existing cloud platforms use traditional GPU configuration specifi cations,which unable to meet the highperformance computing power and GPUs interconnect bandwidth requirements required for large-scale model training.This article mainly studies the intelligent computing chip technology of buckle module form and the cooling technology of the whole computing server,and provides a selection basis for subsequent construction of the intelligent computing center.