Collaborative Research: SHF: Medium: Revitalizing EDA from a Machine Learning Perspective

尽管电子电路和系统的设计自动化在过去取得了巨大的成功，但其有效性和效率仍然有限。这通常是由于点软件工具的不必要的过度迭代，其中对下游设计步骤的早期预测过于悲观，并且不同工具之间的互操作在很大程度上需要手动处理。因此，现有的芯片设计流程不被认为是完全自动化的，并且仍然存在对联合探索这些流程中的不同步骤之间的相当大的空间的强烈需求。此外，现有的设计验证方法通常涉及不必要的冗余和大量的手动工作，大大有助于一个众所周知的瓶颈上市时间。机器学习技术的最新进展为从另一个角度振兴当前的电子设计自动化（EDA）流程提供了一个很好的机会，即，从现有设计数据中提取设计和验证知识，并在新设计中重用。本研究的目标是开发这样的知识提取和重用技术的帮助下，国家的最先进的机器学习技术。这项研究的结果有助于缓解芯片设计生产力危机，并满足对硬件加速计算日益增长的需求。该研究还为包括女性和少数族裔在内的学生提供跨学科技能培训，并为美国未来的高科技劳动力做好准备，以应对电子行业的挑战。该项目涉及在电子设计自动化背景下对机器学习的系统研究，包括五个集成组件：1）开发基于学习的快速和高保真预测技术，用于电路设计的结构和行为领域的知识提取; 2）研究如何将设计预测与电路优化无缝集成：3）将机器学习预测应用于加速功能验证覆盖率和促进自动化调试; 4）发展关于各种工具之间相互作用的自主学习，从而实现自动化综合空间探索; 5）自动化机器-该奖项反映了NSF的法定使命，并被认为值得通过使用基金会的学术价值和更广泛的影响审查标准。

Despite its spectacular success in the past, design automation of electronic circuits and systems remains limited in effectiveness and efficiency. This is often due to unnecessarily excessive iterations of point software tools, where early predictions on downstream design steps are overly pessimistic and interoperations among different tools largely require manual handling. As such, existing chip-design flows are not considered fully automated, and there still exists a strong need for jointly exploring the considerable room between the different steps in these flows. Moreover, existing design-verification approaches usually involve unwanted redundancy and substantial manual effort, contributing greatly to a well-known bottleneck of time-to-market. The recent progress in machine-learning technology offers a great opportunity to revitalize current Electronic Design Automation (EDA) flows from an alternative perspective, i.e., extracting design and verification knowledge from existing design data, and reusing it on new designs. The goal of this research is to develop such knowledge extraction and reuse techniques with the aid of the state-of-the-art machine learning technology. The outcome of this research is to help mitigate the chip-design productivity crisis and cater to the increasing demand for hardware-accelerated computing. This research is also training students, including women and under-represented minorities, with interdisciplinary skills and preparing tomorrow’s high-tech workforce in the U.S. for solving challenges in the electronic industry.The project involves systematic research on machine learning in the context of electronic design automation with five integrated components: 1) development of learning-based fast and high fidelity prediction techniques for knowledge extraction in the structural and behavioral domains of circuit designs; 2) a study on how to seamlessly integrate the design predictions with circuit optimizations; 3) applying machine-learning prediction to accelerating functional-verification coverage and facilitating automated debugging; 4) developing autonomous learning on the interplay amongst tools and thereby achieving automated synthesis space exploration; 5) automated machine-learning architecture search and feature refinement in EDA applications.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.