基于演化计算的处理器结构优化关键技术研究

An important issue in the design flow of a processor is to decide an appropriate processor architecture which optimizes the performance/power of the processor (architecture optimization for short). In the traditional industrial flow, the processor architecture is decided by senior researchers according to their personal experiences as well as extremely-slow processor simulation. In order to enhance the efficiency and reliability of architecture optimization, recently there have been a few studies which use machine learning techniques to model processor performance/power. While helpful in enhancing the efficiency of designing processor, such techniques fail to adapt to three emerging challenges, namely, the heterogeneity challenge, the low-power challenge, and the scalability challenge. Targeting at the three challenges, this project studies a novel methodology of architecture optimization based on evolutionary computation. Inspired by classical paradigms in evolutionary computation, such as large-scale evolutionary optimization and evolutionary game theory, this project proposes a series of key techniques, including evolutionary optimization of heterogeneous multicore architecture, low-power architecture based on evolutionary game theory, and encoding and optimization of scalable interconnection architecture. The ultimate goal is to implement an open-source architecture optimization platform for the research and development of China's domestic processors.

处理器研发中的首要问题是确定合适的处理器体系结构，优化处理器的性能/功耗（以下简称处理器结构优化）。在工业界传统的研发流程中，处理器结构由资深研发人员通过极为缓慢的处理器模拟并结合个人经验来确定。为提高结构优化的效率和可靠性，近年来许多研究者将机器学习方法应用于处理器性能/功耗建模。这类方法对提高处理器研发的效率有很大帮助，但还不能适应处理器异构化、低功耗和可扩展三大新挑战。 本项目面向这三大挑战，研究基于演化计算的处理器结构优化方法。项目组借鉴高维演化优化、演化博弈等演化计算领域的经典学术思想，提出异构多核结构的演化优化、基于演化博弈的低功耗结构、可扩展互连结构的编码及优化等一系列关键技术，力争实现一个处理器结构优化的开源平台，服务于国产处理器的研发。

本项目面向处理器研发中异构化、低功耗以及可扩展三大挑战，研究基于演化计算的处理器结构优化方法。项目组借鉴高维演化优化、演化博弈等演化计算领域的经典学术思想，提出面向异构多核处理器的结构优化、基于演化博弈的低功耗及可扩展片上互连结构优化等一系列关键技术，服务于深度学习处理器的研发。在论文发表方面，项目组在ASPLOS、ISCA、MICRO等会议和ACM、IEEE Trans等期刊上累计发表论文19篇，其中CCF A类期刊/会议论文12篇（MICRO 2018、MICRO 2016、ISCA 2016、MICRO 2015、ISCA 2015、ASPLOS 2015、2TOCS、1TPDS、1TOC、2TCAD）