Maia: A Reticle-Scale AI Accelerator

📋 论文概要

本文介绍了微软的第二代定制AI加速器Maia 200，采用晶圆级集成技术，实现了超过3倍于前代的HBM带宽和计算吞吐量，面向大语言模型工作负载。论文详细阐述了在功率传输、热管理、物理设计、验证和封装方面的创新，提供了一种可扩展的高性能AI加速方案。

💡 主要创新点

• 采用晶圆级（reticle-scale）集成与先进封装技术，实现750W高功率AI SoC的可靠运行。
• 创新功率传输和热管理方案，解决大尺寸芯片供电与散热难题，支持高密度计算。
• 针对大语言模型工作负载进行架构优化，显著提升HBM带宽和计算吞吐量。

🏷 关键词

📄 原文摘要

Microsoft, Mountain View, CA, 2Microsoft, Bengaluru, India 1 Abstract In Paper 17.4, the architecture and implementation of Microsoft’s MAIA AI silicon, a reticlescale 750W AI SoC, is presented. Innovation across power delivery, thermal management, physical design, validation, and packaging is detailed, resulting in a scalable approach to high-performance AI acceleration. Introduction: Maia 200 is Microsoft’s second-generation custom AI accelerator, engineered for large language model workloads. Compared to Maia 100, Maia 200 achieves over 3× higher HBM bandwidth and compute throughput, while scaling network bandwidth by more than 2× (see Fig. 17.4.1 for detailed specifications). Leveraging a vertically integrated approach across the system, networking, and software, Maia 200 is co-designed to deliver leading performance-per-dollar for inference and synthetic data generation workloads. and BGA landing were optimized to meet the target impedance as close as possible. The

👥 作者与机构

S. Xu1, G. Mandal1, S. Jahagirdar1, S. Tripathy2, P. Parthasarathy1, S. Srinivasan1, A. Levin1