Machine learning techniques for taming the complexity of modern hardware design
Machine learning techniques for taming the complexity of modern hardware design
复制标题
用于降低现代硬件设计复杂性的机器学习技术
DOI：
10.1147/jrd.2017.2721699
复制
发表时间：
2017
期刊：
影响因子：
--
通讯作者：
P. Bose
中科院分区：
文献类型：
--
作者：
M. Ziegler;Ramon Bertran Monfort;A. Buyuktosunoglu;P. Bose
对提高新一代IBM服务器性能和效率的不断追求导致了系统复杂性的相应增加。随着硬件复杂性的增加，即需要更多设计选择的更复杂的硬件体系结构，自动化的复杂程度也会增加，以管理设计挑战。现代硬件设计中设计选择的数量要求使用智能自动化技术来导航设计空间。本文介绍了在IBM系统的设计和生命周期中使用的三种基于机器学习的自动化技术。特别是，我们描述了将这些技术应用于IBM z13大型机。在预硅设计阶段，使用一个称为合成调谐系统的软件系统来优化对硬件实现至关重要的合成程序的参数。在设计的前硅阶段和后硅阶段，名为MicroProbe的框架自动生成微基准，即小程序，以确定系统的功率、性能和弹性特性。在客户环境中部署系统产品后，Call Home设施会监控和分析各种现场使用指标，以帮助管理员了解当前系统行为并改进未来的设计。除了现有的IBM系统贡献，这篇高级概述文章还描述了硬件设计领域中的其他机器学习(和相关)技术，以及此类工作的未来方向。
The continual quest to improve performance and efficiency for new generations of IBM servers leads to a corresponding increase in system complexity. As hardware complexity increases, i.e., more complicated hardware architectures requiring more design choices, the level of sophistication in automation also increases to manage the design challenges. The number of design choices in modern hardware design calls for intelligent automated techniques to navigate the design space. This paper covers three machine learning-based automation techniques used during the design and lifetime of IBM systems. In particular, we describe applying these techniques to the IBM z13 mainframe. During the presilicon design phase, a software system called synthesis tuning system is employed to optimize the parameters of the synthesis program vital to hardware implementation. During both the presilicon and postsilicon phases of the design, a framework called MicroProbe automatically generates microbenchmarks, i.e., small programs, to determine power, performance, and resilience characteristics of the system. Following system product deployment in customer environments, the Call Home facility monitors and analyzes a wide range of in-field usage metrics to help administrators understand current system behavior and improve future designs. Beyond existing IBM system contributions, this high-level overview paper also describes additional machine learning (and related) techniques in the field of hardware design, along with future directions for such work.