AMD Instinct MI350 Series GPUs: CDNA 4-Based 3D-Stacked 3nm XCDs and 6nm IODs for AI applications

📋 论文概要

AMD发布了Instinct MI350系列GPU，采用第四代CDNA架构，通过3D堆叠技术将3nm XCD和6nm IOD集成在一起，并利用Infinity Fabric互连和12层HBM3E内存，面向AI应用提供高性能计算能力。

💡 主要创新点

• 采用3D堆叠技术将八颗3nm加速计算芯片（XCD）与两颗6nm I/O芯片（IOD）集成，实现高带宽低延迟互连。
• 引入第四代CDNA架构，针对AI计算优化，支持多达八个12层堆叠的HBM3E内存，显著提升内存带宽和容量。
• 通过AMD Infinity Fabric实现芯片间高速互连，增强可扩展性和能效。

🏷 关键词

📄 原文摘要

Sriram Sundaram1, Mark Silla1, Duncan Law5, Kathy Hoover1, Samuel Lipson1, Kevin Duda2, Vinay Parthasarathy6, Deepesh John1, Hanish Vemulapalli1, Srinivas Pavan Kumar Gade7 AMD, Austin, TX, 2AMD, Fort Collins, CO, 3AMD, Folsom, CA, 4AMD, Shanghai, China, 5AMD, Markham, Canada, 6AMD, San Diego, CA, 7AMD, Hyderabad, India 1 Abstract The AMD Instinct MI350 Series GPU, based on a 4th-generation AMD CDNA architecture, features eight stacked accelerator complex dies (XCD), two I/O dies (IOD) interconnected with AMD Infinity Fabric, and interfaces for eight stacks of 12-hi HBM3E memory. The MI350 Series GPU delivers over 3× generational inference performance increase by improving energy efficiency, increasing compute throughput, supporting new lower-precision FP6/FP4 datatypes, and increasing both memory bandwidth and capacity. The AMD Instinct™ MI350 Series GPUs establish a high level of performance for generative

👥 作者与机构

Ramasamy Adaikkalavan1, Alan Smith1, Teja Singh1, Sundar Rangarajan1, Eric Chapman1, Samuel Naffziger2, Subramaniam Maiyuran3, Candy Chen4,