A Software-Assisted Peak Current Regulation Scheme to Improve Power-Limited Inference Performance in a 5nm AI SoC

📋 论文概要

本文提出了一种软件辅助的峰值电流调节方案，通过动态管理AI SoC的峰值电流，在功率受限条件下提升推理性能。该方案在5nm工艺上实现，有效解决了高功耗导致的性能瓶颈问题。

💡 主要创新点

• 提出软件辅助的峰值电流动态调节机制，结合硬件协同优化，实现低延迟的功率管理。
• 针对AI推理工作负载的特性，设计峰值电流感知的调度策略，显著提升功率受限时的吞吐量。
• 在5nm先进工艺节点上实现，验证了软硬件联合优化的有效性，为高性能AI芯片设计提供新思路。

🏷 关键词

📄 原文摘要

Bruce Fleischer1, Joshua Rubin1, JohnDavid Lancaster1, Saekyu Lee1, Matthew Cohen1, Matthew Ziegler1, Nianzheng Cao1, Sandra Woodward2, Ankur Agrawal1, Ching Zhou1, Prasanth Chatarasi1, Thomas Gooding2, Michael Guillorn1, Bahman Hekmatshoartabari1, Philip Jacob1, Radhika Jain1, Shubham Jain1, Jinwook Jung1, Kyu-Hyoun Kim1, Siyu Koswatta1, Martin Lutz1, Alberto Mannari3, Abey Mathew4, Indira Nair1, Ashish Ranjan1, Zhibin Ren1, Scot Rider5, Thomas Roewer1, David Satterfield6, Marcel Schaal1, Sanchari Sen1, Gustavo Tellez1, Hung Tran1, Wei Wang1, Vidhi Zalani1, Jintao Zhang1, Xin Zhang1, Vinay Shah7, Robert Senger1, Arvind Kumar1, Pong-Fei Lu1, Leland Chang1 IBM Research, Yorktown Heights, NY; 2IBM, Rochester, MN IBM Research, Zurich, Switzerland; 4IBM, Austin, TX; 5IBM, Poughkeepsie, NY 6 IBM Research, Lowell, MA; 7IBM, Hursley, United Kingdom 1 3 The rapid emergence of AI models, specifically large language models (LLMs) requiring

👥 作者与机构

Monodeep Kar1, Joel Silberman1, Swagath Venkataramani1, Viji Srinivasan1,