Design Automation and Architecture Design for Near Threshold Computing

物联网(IoT)是一个快速增长的市场，对工业和个人生活的各个方面都产生了重大影响。物联网边缘设备严重依赖超低功耗和更小的外形尺寸。在各种使用场景中，许多物联网设备可以收集能量或使用有限的电池，预计将运行很长的运行寿命。它们还要求适应动态能源，并调整性能、准确性和正确性，以满足严格的能源预算。在这种背景下，近阈值计算(NTC)通过将电源电压降低到晶体管的阈值电压，是一种很有前途的显著降低能耗的方法。然而，在NTC电路、体系结构和系统设计方面存在各种挑战，这需要全新的分析框架、电子设计自动化(EDA)工具和设计范式。特别是，由于更高的故障率和对变化的敏感性，降低电源电压是以显著的弹性损失为代价的。因此，能效恢复能力是在NTC运行的物联网边缘设备面临的主要挑战。该方案的主要前提是为近门限计算系统开发跨层可靠的设计自动化和体系结构设计方法，这将推动物联网的许多新兴应用领域。通过从电路、逻辑设计和自动化流程到NTC系统架构的一系列跨层技术，实现了弹性NTC电路和体系结构的正确设计。我们特别提出发展i)逻辑和内存优化以在接近阈值状态下运行可靠性，ii)运行时调优和自适应以预测和减轻由于接近阈值操作而导致的故障，以及iii)通过近似计算包含接近阈值计算中的残差。

Internet of Things (IoT) is a fast growing market with significant impacts in various aspects of industry and personal life. IoT edge devices critically depend on ultra-low power and reduced form-factor. Many IoT devices in various usage scenarios could be energy-harvested or work with limited battery, and expected to run for a long operational lifetime. They also require to adapt for energy on-the-fly, and adjust performance, accuracy, and correctness to meet the stringent energy budget. In this context, near threshold computing (NTC), by lowering the supply voltage down to the threshold voltage of transistors, is a promising approach to significantly reduce energy consumption. However, there are various challenges in NTC circuit, architecture and system design, which require completely new analysis frameworks, electronic design automation (EDA) tools, and design paradigms. In particular, reducing the supply voltage comes at the expense of significant resiliency loss, due to higher failure rates and increased sensitivity to variations. Therefore, energy-efficient resiliency is a major challenge for IoT edge devices operating at NTC. The main premise of this proposal is to develop cross-layer reliable design automation and architecture design methodologies for near threshold computing systems, which drive many emerging application domains in IoT. Proper design of resilient NTC circuits and architectures is achieved by a series of cross-layer techniques from circuit, logic design and automation flows all the way to the system architecture for NTC. We specifically proposed to develop i) logic and memory optimization to operate reliability at near threshold regime, ii) runtime tuning and adaptation to predict and mitigate failures due to near threshold operation, and iii) embracing residual errors in near threshold computing through approximate computing.