Innovation For the Next Decade of Compute Efficiency

📋 论文概要

本文展望了未来十年计算效率的创新方向，讨论了高性能计算面临的挑战，并提出了通过架构创新、先进封装和软件协同设计来持续提升计算性能的路径。

💡 主要创新点

• 提出通过异构计算和专用加速器来突破通用处理器的能效瓶颈
• 强调先进封装技术（如Chiplet、3D堆叠）在提升计算密度和带宽中的作用
• 倡导软硬件协同设计，利用领域特定架构优化数据流和功耗

🏷 关键词

📄 原文摘要

AMD, Fort Collins, CO2 1.1 Introduction With high-performance computing becoming an increasingly essential part of modern life, efficiently delivering improvements in compute performance is the defining challenge for our industry. The rapid growth in the number of connected devices is creating ordersof-magnitude more data, requiring greater levels of compute to generate actionable insights for personal use and businesses. At the same time, supercomputers have become critical to enable research breakthroughs in fields including climate change, renewable energy development, infectious disease research, complex life science modeling, and much more. Over the past two decades the industry has consistently doubled performance every 2 to 2.5 years through continued innovation, even as Moore’s Law has slowed down (Figures 1.1.1 and 1.1.2). These increases in compute capability were realized through

👥 作者与机构

Lisa Su1, Sam Naffziger2

AMD, Austin, TX1