智能存算融合系统研究进展与发展趋势

Research Progress and Development Trends of Intelligence Storage-Computation Fusion Systems

随着人工智能与万物互联时代的到来,海量数据驱动的大模型应用场景和计算任务层出不穷,深刻改变了信息技术与装备技术的发展。然而,在传统的冯·诺依曼体系结构计算系统中,存算分离的结构特性引发了“存储墙”问题,限制了计算系统处理计算密集型应用时的能效性能提升。存算融合系统是后摩尔时代下,衔接传统架构与新的类脑计算架构的新型计算范式,利用存储设备与定制化领域专用加速器的特性,将应用中的数据密集型计算任务融入存储设备中,从而消除数据搬运的开销,实现智能计算系统能效性能的显著提升。论文综述了存算融合系统研究进展,并聚焦于智能数据库这一典型基础应用,阐述了存算融合系统的发展趋势与前景。

With the advent of the era of artificial intelligence and the Internet of Things,scenarios and computing tasks driven by massive data and large-scale models are emerging one after another,which has profoundly changed the development of informa⁃tion technology and equipment technology.However,in traditional Von Neumann architecture computing systems,the separatist structure of storage and computation has caused the"memory wall"problem,limiting the energy-efficiency performance improve⁃ment of computing systems when dealing with compute-intensive applications.Storage-Computation Fusion System is a new comput⁃ing paradigm that bridges traditional architecture with new neuromorphic computing architecture in the post-Moore's Law era.By uti⁃lizing the characteristics of storage devices and customized domain-specific accelerators,the system integrates data-intensive com⁃puting tasks in applications into storage devices,thus eliminating the overhead of data movement and achieving significant improve⁃ments in energy efficiency and performance of intelligent computing systems.This paper summarizes the research progress of stor⁃age-computation fusion systems with focusing on intelligent databases,which is a typical fundamental application.It discusses the development trends and prospects of storage-computation fusion systems.