Summit and Sierra: Designing AI/HPC Supercomputers

📋 论文概要

该论文介绍了IBM为美国能源部开发的Summit和Sierra超级计算机系统，分别部署于橡树岭国家实验室和劳伦斯利弗莫尔国家实验室。这些系统旨在提供显著的计算能力提升，是迈向百亿亿次计算的关键一步。

💡 主要创新点

• 采用混合CPU-GPU架构，结合IBM POWER9处理器和NVIDIA Volta GPU，实现高效能AI和HPC工作负载加速。
• 引入了高带宽内存和快速互连网络（NVLink和PCIe 4.0），以缓解数据移动瓶颈，提升整体系统性能。
• 在系统级设计中优化了电源管理和冷却方案，实现了高密度部署下的可靠性及能效平衡。

🏷 关键词

📄 原文摘要

IBM Research, Yorktown Heights, NY 3 IBM Systems and Technology, Austin, TX 1 2 The Summit and Sierra Supercomputer Systems, deployed in 2018 at the Department of Energy (DOE) National Laboratories, Oak Ridge (ORNL) and Lawrence Livermore (LLNL), provide a significant increase in computing capability relative to the predecessors, and represent a major step in the path to Exascale computing. Arising from the DOE CORAL program (https://asc.llnl.gov/CORAL/), Summit and Sierra were listed as #1 and #2, respectively, in the November 2018 version of the Top500 list (https://www.top500.org/lists/2018/11/). Supercomputers such as Sierra and Summit have unique design challenges in network bandwidth, communication patterns, synchronous coordination and parallel computation not found in traditional cloud deployments. While Supercomputers and Data Centers are both large collections of interconnected compute nodes

👥 作者与机构

James A. Kahle1, Jaime Moreno2, Dan Dreps3

IBM Research, Austin, TX