差错容忍的近阈值计算技术研究

Near-threshold computing is one of the effective means to improve energy efficiency. However, the reduction of voltage makes the chips face severe reliability issues, and providing accurate and reliable computing is often costly. Approximate computing opens the door for high-reliability and low-power design for systems that allow tolerable error rates. Unfortunately, the near-threshold voltage operations lead to significant variations in processor cores’ threshold voltage and failure rate between the cores, and the quality of approximate computing across different applications is also different. It is crucial to solve these challenges and ensure that the systems deliver optimal performance/power efficiency, dependability and user experience. The proposed research consists of the following four tasks: (1) studying the approximate computing quality model based on NTC failure rate, which will provide theoretical foundation for this research; (2) investigating approximate computing quality-aware power management methods, which aim to solve threshold voltage variations issues among different cores; (3) exploring thermal-aware application scheduling schemes, which help alleviating the hot-spots induced reliability issues; and (4) experimenting hardware and software co-design of a multi-core system in the context of near threshold voltage computing, which will provide design space exploration and verification platform for this study. If successful, the proposed research will significantly improve the reliability of near-threshold voltage computing technology. More importantly, it will ensure the continuous scaling of multi-core computing density while achieving the goals of delivering high-performance, low-power and high-reliability. It is expected that 7-10 high-quality research papers will be published via professional conferences or journals and 3-5 patent applications will be filed.

近阈值计算是提高计算机系统能效性的有效手段之一，但电压降低却导致了严峻的可靠性问题。提供精确可靠的计算往往开销较高，近似计算技术通过容忍差错为高可靠、低功耗的系统设计提供了可能。然而不同应用的近似计算质量存在差异性，近阈值电压又加剧了处理器核间的差异性。如何解决上述差异性是优化系统能效性、可靠性和用户体验等多方面的核心问题。因此，课题拟深入探索如下四个方面的内容：(1)研究基于NTC失效率的近似计算质量模型，为系统动态管理和优化提供理论支持；(2)针对近阈值电压引发的差异化问题，研究计算质量感知的功率管理方法；(3)针对热点影响可靠性的问题，研究温度感知的应用程序调度方案；(4)研究支持近似计算的近阈值系统软硬件协同设计技术。本课题将极大改善近阈值电压技术面临的可靠性问题，实现性能、功耗和可靠性的综合优化。课题预期在国内外期刊和国际会议上发表7-10篇高水平论文，并申请3-5项专利。

近阈值计算是提高能效性的有效手段之一，但使得芯片面临了严峻的可靠性问题。提供精确可靠的计算往往开销较高，近似计算技术通过容忍差错为高可靠、低功耗的系统设计提供了可能。然而不同程序的近似计算质量存在差异，近阈值电压更加剧了处理器核间的差异性。如何解决上述差异使得软硬件匹配是优化系统能效性、可靠性和用户体验等多目标的核心问题。课题的主要贡献包括以下几个方面：(1)研究基于NTC失效率的近似计算质量模型，为系统动态管理和优化提供理论支持；(2)针对近阈值电压引发的差异化问题，研究计算质量感知的功率管理方法；(3)针对软错误导致的可靠性问题，研究故障感知的应用程度调度和裁剪配置方案；(4)研究支持近似计算的近阈值系统软硬件协同设计技术。本课题极大改善近阈值电压技术面临的可靠性问题，实现了性能、功耗和可靠性的综合优化。课题发表国内外期刊和国际会议论文12篇，并申请专利3项。