Enabling Adaptive Voltage Regulation: Control, Machine Learning, and Circuit Design

电源电压调节在以良好调节的电压电平向片上器件供电方面发挥着关键作用。电压调节是从高性能微处理器到移动的片上系统的电子系统的关键设计挑战。在这样的系统中，必须满足对处理能力不断增长的需求，同时保持在指定的功率、热量和电池寿命限制内。 必须在以各种可能的方式最大化系统电源效率的同时管理和提供电源。拟议的研究旨在通过采取跨学科的方法来解决上述电压调节挑战。 将开发控制、机器学习和电路设计方面的创新，以实现涉及各种片上/片外电压调节器的自适应电源电压调节系统。该项目的预期成果将有助于构建新一代高效电路和系统，这些电路和系统可以自适应变化的操作条件。该项目所追求的电路/系统设计，控制理论探索和机器学习之间的协同作用将促进电子系统设计的新的跨学科方向。 这项研究的深度和广度将使学生获得出色的教育和培训机会。本科生和代表性不足的学生的参与是该项目的一项重要教育使命，将通过招募和外展来促进。该项目的预期成果预计将是广泛的，并将广泛传播，以及带到课堂上，以有利于本科生和研究生课程。 与产业界的合作与互动，是本项目影响真实的世界的重要渠道，将积极推进。 该项目基于这样的愿景，即通过从板上开关稳压器（VR）开始的异构电压处理链，到封装内/片上开关VR，最后到分布式片上线性VR网络，可以最好地实现电源电压调节的最终质量和效率。 异构电压调节（HVR）系统是有前途的，因为它们包括在响应时间、尺寸、效率和成本方面具有互补折衷的调节器。该项目的最终目标是使HVR系统能够保证电源完整性，产生最小的功率损耗，并在多个时间尺度上自主适应工作负载变化和系统/环境不确定性。 上述目标将通过寻求新型控制理论、电路和机器学习支持的自主适应的集成解决方案来实现。将为分布式片上线性调节器网络和具有保证稳定性和调节性能的HVR系统开发分散和集中控制的严格设计技术。 高效的机器学习算法及其片上集成将用于提供时变负载电流的准确实时预测。HVR系统的自主自适应将得到高能效控制策略的支持，这些策略基于机器学习预测的未来电流负载，抢先适应片上线性调节器网络和片上/片外VR。 应对系统不确定性是另一个关键目标，将通过部署控制策略来实现，这些控制策略通过机器学习进行自我调整，以实现最佳的功率效率。 该项目将探索系统级设计优化，以联合优化HVR系统中所有电压处理阶段的调节性能、功率效率和设计开销。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Supply voltage regulation serves the critical role of delivering power to on-chip devices at well-regulated voltage levels. Voltage regulation presents key design challenges of electronic systems ranging from high-performance microprocessors to mobile system-on-a-chips. In such systems, the ever-growing need for processing capability must be fulfilled while staying within specified power, thermal, and battery-life limits. Power must be managed and delivered while maximizing system power efficiency in every possible way. The proposed research aims to address the above voltage regulation challenges by taking an interdisciplinary approach. Innovations in control, machine learning, and circuit design will be developed to enable adaptive supply voltage regulation systems involving a variety of on-chip/off-chip voltage regulators. The expected outcomes of this project will help build new generations of highly efficient circuits and systems that can self-adapt to varying operating conditions. The synergies between circuit/system design, control-theoretical exploration, and machine learning as pursued in this project will promote a new interdisciplinary direction for advancing electronic system design. The depth and breadth of this research will expose students to outstanding educational and training opportunities. Participation from undergraduate and underrepresented students is an important education mission of this project and will be promoted through recruiting and outreaching. The anticipated results from this project are expected to be broad and will be widely disseminated as well as brought to classroom to benefit undergraduate and graduate curriculum. Collaboration and interaction with industry constitutes an important channel for this project to impact the real world, which will be actively pursued. This project is based on the vision that the ultimate quality and efficiency in supply voltage regulation may be best achieved via a heterogeneous chain of voltage processing starting from on-board switching voltage regulators (VRs), to in-package/on-chip switching VRs, and finally to networks of distributed on-chip linear VRs. Heterogeneous voltage regulation (HVR) systems are promising as they encompass regulators with complimentary tradeoffs in response time, size, efficiency, and cost. The ultimate aim of this project is to enable HVR systems that will guarantee power integrity, incur minimal power loss, and autonomously adapt to workload changes and system/environmental uncertainties at multiple temporal scales. The above goal will be achieved by pursuing an integrated solution of novel control theory, circuits, and machine-learning enabled autonomous adaptation. Rigorous design techniques for decentralized and centralized control will be developed for distributed on-chip linear regulator networks and the HVR system with guaranteed stability and regulation performance. Efficient machine-learning algorithms and their on-chip integration will be employed to provide accurate real-time prediction of time-varying load currents. Autonomous adaptation of the HVR system will be supported by power-efficient control policies that preemptively adapt on-chip linear regulator networks and on-chip/off-chip VRs based on machine-learning predicted future current loads. Coping with system uncertainties is another key objective and will be achieved via deployment of control policies that are self-tuned by machine learning to attain the optimal power efficiency. The project will explore system-level design optimization to jointly optimize regulation performance, power efficiency, and design overhead across all voltage processing stages in a HVR system.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.