Collaborative Research: DESC: Type I: SEEDED: Sustainability-aware Reliable and Reusable AI Hardware Design

深度神经网络(DNN)加速器是专为深度学习算法等计算密集型应用而设计的硬件。随着人工智能(AI)和机器学习在当今许多电子系统中的使用，DNN加速器正迅速成为常见的地方。其中一些应用包括汽车、医疗保健和航空航天等安全关键型应用。然而，许多高精度的DNN加速器消耗大量能源，使得它们在能源受限的设备上的使用受到限制。这个问题的一种流行的解决方案是硬件近似，其中加速器的近似设计被认为足以达到能源效率的目的。使用硬件近似可以使结果比精确的结果更容易受到永久性故障的影响高达3倍。永久性故障通常会导致在制造后阶段报废。当芯片正在使用时，这种永久性故障也可能出现在部署后阶段。这两种情况都是环境成本高昂且不可持续的。这项研究的主要目的是开发一种新的可持续发展意识的设计流程，用于近似深度神经网络-DNN，通过支持重用和自我修复来延长其寿命，同时优化性能和可持续发展指标。在早期设计阶段的可持续发展意识设计流程，并增加具有轻量级故障检测和自修复能力的近似硬件DNN加速器(AxDNN)，使其能够在接近原始基线精度的情况下重复使用和重新利用。在这一目标的推动下，本项目旨在改变AxDNN设计和部署的最新水平，并提出了三个研究目标：(I)借助一种新颖高效的神经结构搜索方法，研究设计可靠和可持续的AxDNN的方法；(Ii)通过旁路电路、近似再训练、混合内建自测试和通过权重交换进行自我修复，研究AxDNN的制造后和部署后故障缓解方法；以及，(Iii)开发模拟和现场可编程门阵列(FGA)演示平台，以评估和演示所产生的AxDNN相对于用户定义的一阶指标和新的可持续发展感知维度的有效性。项目成果(即新的理论、工具、代码、基准和案例研究)将通过开源软件和同行评审出版物向更广泛的机器学习和网络物理系统(CPS)社区公开提供。此外，项目成果将为计算机和电气工程本科生和研究生课程创建新的课程和动手实验室练习。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Deep Neural Network (DNN) accelerators are hardware designed for compute-intensive applications like deep learning algorithms. DNN accelerators are quickly becoming common place as artificial intelligence (AI) and machine learning are used in many electronic systems today. Some of these include safety-critical applications such as automotive, healthcare, and aerospace. However, many of the high-precision DNN accelerators consume a lot of energy, making their use limited for energy-constrained devices. A popular solution to this issue has been hardware approximation, where an approximate design of the accelerator is considered to be sufficient for purposes of energy efficiency. The use of hardware approximation can make the outcome up to 3X more vulnerable to permanent faults compared to their accurate counterparts. Permanent faults usually lead to discarding in the post-fabrication phase. Such permanent faults can also appear in the post-deployment phase when the chip is in use. Both cases are environmentally costly and not sustainable. The main objective of this research is to develop a new sustainability-aware design flow for approximate deep neural networks - DNNs that can prolong their lifetime by enabling reuse and self-repair while also optimizing performance and sustainability metrics.A sustainability-aware design flow in the early design phases and augmenting approximate hardware-based DNN accelerators (AxDNNs) with lightweight fault detection and self-repair capability allows their reuse and re-purpose with operating capability close to the original baseline accuracy. Motivated by this goal, this project aims to transform the state-of-the-art in designing and deploying AxDNNs with three proposed research objectives: (i) investigating methods for designing reliable and sustainable AxDNNs with the help of a novel and efficient neural architecture search methods; (ii) investigating methods for post-fabrication and post-deployment fault mitigation in AxDNNs with the help of bypass circuitry, approximate retraining, hybrid built-in-self-test, and self-repair through weight swapping; and, (iii) developing a simulation and field programmable gate arrays (FPGAs) demonstration platform to evaluate and demonstrate the effectiveness of the resulting AxDNNs compared against user-defined first-order metrics and novel sustainability-aware dimensions. The project outcomes (i.e., new theories, tools, codes, benchmarks, and case studies) will be publicly made available to the broader machine learning and cyber-physical system (CPS) communities through open-source software and peer-reviewed publications. In addition, the project outcomes will create a new curriculum and hands-on laboratory exercises for computer and electrical engineering undergraduate and graduate courses.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.