SHF: Medium: ASKS - Architecture Support for darK Silicon

拟议的ASKS(暗硅体系结构支持)项目建议对未来采用新兴技术设计的多核微处理器提供体系结构支持，以应对即将到来的暗硅革命的挑战。为了保持历史性能曲线，未来的多核处理器包含的晶体管将超过所有芯片同时可持续供电的能力。因此，我们面临着？暗硅？的前景。其中有一些芯片？S的组件必须断电(变暗)才能留在芯片内？S的电源/散热预算。在这些情况下确保最佳性能是一项具有挑战性的任务。具体地说，在具有不同类型组件(核心、高速缓存、片上网络(NOC或片上互连)、存储器控制器等)的体系结构中，在功率预算内可以有许多不同的暗淡配置。然而，它们可能表现出显著不同的性能。此外，非核心组件(共享缓存、片上互连、存储体系结构)的功耗很大，因此非核心组件在联合性能/功耗/热优化中发挥着重要作用。因此，需要一种更具包容性的方法。我们的架构设计空间探索将主要集中在非核心空间(共享缓存、片上互连结构、存储器体系结构)，目标是未来基于NoC的多核微处理器，这些微处理器利用新兴的3D芯片堆叠(3D IC)和非易失性存储器(NVM)技术。该项目将在两个主要方面推进为黑暗硅革命做准备的最先进水平。1)跨层整体优化，关注非核心部件：非核心部件在联合性能/功率/散热优化中起着重要作用。该项目提出了一种集成的方法，在该方法中，核心和非核心组件在功率和热量限制以及动态变化的程序行为和执行参数的情况下协作以最大化性能。2)在新兴技术的背景下研究暗硅：新兴的3D IC和NVM技术被认为是设计未来多核体系结构的有前途的方法。这些新兴技术的采用对暗硅提出了新的挑战(例如3D堆叠芯片中的热分布加剧)，但也为架构创新带来了新的机遇，例如新颖的电源管理技术和更大程度地利用存储系统异构性。该项目的更广泛影响包括对未来微处理器性能提高的贡献，即使面对即将到来的黑暗硅革命。通过与几个行业合作伙伴的密切合作，PI设想将许多想法直接转移到行业中。本项目开发的工具和技术将用于教授现有课程和开发新课程，并将通过网络提供给其他教育工作者、研究人员和行业从业者使用。研究成果的传播还将通过会议教程、小组讨论和讲习班进行。我们将共同努力，让代表不足的群体和本科生参与到这项研究中来。

The proposed ASKS (Architecture Support for darK Silicon) project proposes architectural support for future many-core microprocessors designed with and around emerging technologies to address the challenges of the coming dark silicon revolution. Future many-cores, to stay on the historical performance curve, will contain more transistors than can all be sustainably powered at the same time. Thus, we face the prospect of ?dark silicon,? wherein some of a chip?s components must be powered-off (darkened) in order to stay within the chip?s power/thermal budget. Ensuring peak performance under these circumstances is a challenging task. In particular, in an architecture with different types of components (cores, caches, network on chip (NoC, or on-chip interconnet), memory controllers, etc.), there can be many different on-dim-dark configurations within the power budget. However, these can exhibit significantly varying performances. Additionally, the power consumption of uncore components (shared caches, on-chip interconnect, memory architecture) is significant, thus the uncore components play an important role in joint performance/power/thermal optimization. Therefore, a more inclusive approach is needed. Our architectural design space exploration will mainly focus on the uncore space (shared caches, on-chip interconnect structures, memory architecture), targeting future NoC-based many-core microprocessors that exploit the emerging technologies of 3D die-stacking (3D IC) and non-volatile memories (NVM). This project will advance the state of the art in preparing for the dark silicon revolution in two main aspects. 1) Cross-layer holistic optimization with a focus on uncore components: The uncore components play important roles in joint performance/power/thermal optimization. This project proposes an integrated approach in which the cores and uncore components collaborate to maximize performance under power and thermal constraints as well as under dynamically changing program behavior and execution parameters. 2) Investigating dark silicon in the context of emerging technologies: Emerging 3D ICs and NVM technologies are envisioned as promising ways to design future many-core architectures. The adoption of such emerging technologies poses new challenges for dark silicon (such as aggravated thermal profiles in 3D stacked chips), but also brings new opportunities for architectural innovations such as novel power management techniques and greater exploitation of memory system heterogeneity. The broader impact of this project includes the contribution to the increased performance for future microprocessors even in the face of the coming dark silicon revolution. Through close collaboration with several industry partners, the PIs envision direct transfer of many ideas to industry. The tools and techniques developed in this project will be used in teaching existing courses and developing new courses, and will be made available through the web for use by other educators, researchers, and industry practitioners. Dissemination of research findings will also be carried out through conference tutorials, panel discussions, and workshops. A concerted effort will be made to involve under-represented groups and undergraduate students in this research.