摘要
阐述互联技术与人工智能芯片在大模型中的特点。介绍大模型中常用的并行拆分方式,提出一种适合不同规模张量并行的物理拓扑,并针对不同规模的应用,分析对应的逻辑拓扑和数据处理方法,从而高效地满足大模型对互联的需求。
This paper expounds the characteristics of interconnection technology and artificial intelligence chips in large models.It introduces the commonly used parallel splitting methods in large models,proposes a physical topology suitable for parallel tensors of different scales,and analyzes the corresponding logical topology and data processing methods for applications of different scales,thereby efficiently meeting the interconnection requirements of large models.
作者
赵立东
孟玉
刘彦
ZHAO Lidong;MENG Yu;LIU Yan(Shanghai Enflame Technology Co.,Ltd.,Shanghai 201306,China)
出处
《集成电路应用》
2024年第6期22-24,共3页
Application of IC
关键词
人工智能芯片
互联技术
分布式系统
artificial intelligence chips
interconnection technology
distributed system