信息通信技术与政策

信息通信技术与政策

信息通信技术与政策 ›› 2024, Vol. 50 ›› Issue (6): 2-9.doi: 10.12267/j.issn.2096-5931.2024.06.001

专题:先进计算 上一篇    下一篇

大模型算力基础设施技术趋势、关键挑战与发展路径

Large model computing infrastructure technological trends, key challenges, and development trajectories

张政, 冯少飞   

  1. 浪潮电子信息产业股份有限公司,北京 100089
  • 收稿日期:2024-05-10 出版日期:2024-06-25 发布日期:2024-07-30
  • 作者简介:
    张政, 浪潮电子信息产业股份有限公司产品总监,计算机科学与技术博士后,主要从事计算机体系结构、异构计算等方面技术研究和产业化工作;
    冯少飞, 浪潮电子信息产业股份有限公司高级产品经理,计算机科学与技术博士后,主要从事计算机体系结构、绿色计算方面技术研究和产业化工作

ZHANG Zheng, FENG Shaofei   

  1. IEIT SYSTEMS Co., Ltd., Beijng 100089, China
  • Received:2024-05-10 Online:2024-06-25 Published:2024-07-30

摘要:

从大模型技术发展趋势出发,分析了多模态、长序列和混合专家模型的架构特征和算力需求特点。围绕大模型对巨量算力规模与复杂通信模式的需求,重点从算力利用效率、集群互联技术两方面量化分析了当前大模型算力基础设施存在的发展问题和面临的技术挑战,并提出了以应用为导向、以系统为核心、以效率为目标的高质量算力基础设施发展路径。

关键词: 多模态模型, 长序列模型, 混合专家模型, 算力利用效率, 集群互联, 高质量算力

Abstract:

Starting from the latest technological development trends of large models, this paper first analyzes the architectural characteristics and computing power demand features of multimodal, long sequence, and mixture of experts models. Further, it focuses on the requirements of the latest large models for massive computing power scale and complex communication patterns. It quantitatively analyzes the current development problems and technical challenges faced by large model computing infrastructure from two aspects: computating efficiency and cluster interconnection technology. Finally, it proposes a high-quality computing infrastructure development trajectory oriented by applications, centered on systems, and targeted at efficiency.

Key words: multimodal model, long sequence model, mixture of experts model, computating efficiency, cluster interconnection, high-quality computing power

中图分类号: