Combining Testing and Monitoring for Online Functional Guarantees in Imprecise Hardware Systems

两股力量正在塑造新数字系统的设计：可变性和近似性。随着技术规模的扩大，器件和互连的可变性由于固有的（例如，不相等的掺杂剂浓度）和非本征（例如，温度变化）因素。此外，设备会随着时间的推移而退化-所谓的老化-这加剧了可变性问题。为了缓解这些问题，提出了新的设计方法，在设计的某些部分和某些时间段内接受“不精确”的功能。因此，错误正在成为设计流程的一个组成部分。“不精确”功能以两种根本不同的风格出现：要么是由于技术的不可避免的可变性，这是通过额外的逻辑来解决的-所谓的随机处理-或者是通过设计来减少系统的面积或能耗-所谓的近似处理。在考虑未来的系统时，这两个特性将发挥重要作用，例如，目前，当不精确处理系统包括易受老化和环境变化影响的不精确硬件时，不可能在运行时可靠地量化不精确处理系统的准确性保证。因此，安全地部署使用函数近似并在新的易变技术中实现的可适应系统是困难的。我们的主要目标是克服这个问题，开发新的概念，为不精确的电路，提供功能保证的第一个运行时诊断基础设施。我们希望通过结合新的算法来实现这一目标，这些算法用于生成和合成基于刺激的测试，并与新的硬件构建块紧密结合，用于在运行时监控系统的功能。我们将在片上视觉系统上验证我们的设计工具和方法。

Two forces are shaping the design of new digital systems: variability and approximation. As technology scales, the variability of devices and interconnects increases dramatically due to intrinsic (e.g., unequal dopant concentrations) and extrinsic (e.g., temperature variations) factors. Furthermore, devices degrade over time -the so-called aging- which exacerbates the variability problem. To palliate these issues, new design methodologies proposed to accept an "imprecise" functionality at some parts of the design and during some periods of time. Thus, errors are becoming an integral part of the design flow. The "imprecise" functionality appears in two fundamentally different flavors: either it is due to the unavoidable variability of the technology which is addressed with additional logic -the so-called stochastic processing- or it is created by design to reduce area or energy consumption of the system -the so-called approximate processing. These two characteristics will play an essential role when considering future systems, e.g., enhancing the infrastructure for the Internet of Things (IoT) and creating multiprocessor systems for high-performance computing.Currently, it is not possible to quantify the accuracy-warranties of an imprecise processing system reliably at run time when it includes imprecise hardware subject to aging and environmental changes. Consequently, safely deploying adaptable systems that use functional approximation and that are implemented in new variability-prone technologies is difficult. Our main goal is to overcome this problem by developing new concepts for the first run-time diagnostic infrastructure for imprecise circuits that provides functional guarantees. We want to achieve this by combining new algorithms for generation and synthesis of stimuli-based testing tightly joined with new hardware building blocks for monitoring functionality of a system at run time. We will validate our design tools and methodology on a Vision-System-on-Chip.