AMD InstinctTM MI300 Series Modular Chiplet Package – HPC and AI Accelerator for Exa-Class Systems

📋 论文概要

本文介绍了AMD Instinct MI300系列加速器，通过模块化芯片封装技术首次将数据中心级CPU、GPU、Infinity Cache和8堆栈HBM3内存集成到单个封装中，旨在解决AI和HPC应用中的内存带宽瓶颈，提供超过5TB/s的内存带宽，支持Exa级系统。

💡 主要创新点

• 首次将数据中心级CPU、GPU、Infinity Cache和8堆栈HBM3内存集成到单个封装中，实现高带宽内存访问。

📄 原文摘要

MA 1 4 The AMD InstinctTM MI300 Series accelerators were conceptualized to extract maximum HPC and AI capability from the latest silicon and advanced packaging technology, designed to operate as CPU hosted PCIe® device, MI300X, as well as a self-hosted accelerated processing unit (APU), MI300A. AMD chiplet capabilities and advanced packaging allow AMD’s first-ever integration of data center class CPU, GPU accelerated compute, AMD Infinity Cache, and 8-stack HBM3 memory system into a single package. Observing that many AI and HPC operators are memory bound, AMD targeted MI300 to deliver over 5TBps of HBM3 peak bandwidth. MI300X (Fig. 11.1.8) targets traditional 2P CPU servers hosting eight GPU accelerators with host DDR and device HBM for large model AI training and inference. MI300X

👥 作者与机构

Alan Smith1, Eric Chapman1, Chintan Patel1, Raja Swaminathan1, John Wuu2,

Tyrone Huang3, Wonjun Jung3, Alexander Kaganov3, Hugh McIntyre4, Ramon Mangaser5 AMD, Austin, TX; 2AMD, Fort Collins,CO; 3AMD, Markham, Canada AMD, Santa Clara, CA; 5AMD, Boxborough,