CRII: SHF: Real-time Approximate-Dynamic-Programming based Neuro-controllers for Dynamic Power Management in Power-Constrained Digital Systems

随着物联网（IoT）革命的发展，具有嵌入式能量采集器和集成电池的自供电设备已成为现实。这样的能量源在其电压、电流和功率输出方面易于发生大的变化。同时，负载电路通过细粒度的时空功率管理，增加电源域的数量，减少电压域之间的解耦和不可靠的寄生效应，经历大的动态范围。 然而，传统的电力输送网络（包括DC-DC转换器和电压调节器）仍然是针对最坏情况设计的，因此存在严重的效率低下。这导致整个系统的非最佳功率、性能和能量效率。该项目提出了一种新颖的颠覆性技术，设想在物联网设备中建立一个高度动态的电力输送网络，该网络可以自主适应，重新配置和管理自己，以满足不断变化的电源和负载条件。该研究从近似动态规划的最新进展中汲取灵感，以在高度动态条件下提供电力输送网络的实时和最佳控制。将开发基于硬件的控制器，以提供嵌入式稳压器和DC-DC转换器的实时优化，从而在性能约束下实现最大能效。这种方法的成功取决于电力输送网络的进步，这也将被探讨。传统的？静电干扰？设计不能在飞行中被控制。因此，本研究的第二个主题是探索？变结构控制作为实现超快速和动态可重新配置的功率输送系统的手段。这将在广泛的动态操作范围内实现能源效率的数量级提高，并允许物联网设备的新应用具有深远的社会影响。

With the Internet of Things (IoT) revolution, self-powered devices with embedded energy harvesters and integrated batteries have become a reality. Such energy sources are prone to wide variations in their voltage, current and power outputs. Simultaneously, load circuits undergo large dynamic ranges through fine-grained spatio-temporal power management, increasing number of power domains, decreasing decoupling between voltage domains and unreliable parasitics. However, the traditional power delivery network (which includes DC-DC converters and voltage regulators), are still designed for the worst-case corner and hence suffer from serious inefficiencies. This results in non-optimal power, performance and energy-efficiency of the overall system. This project proposes a novel and disruptive technology, where a highly dynamic power delivery network in IoT devices is envisioned, which can autonomously adapt, reconfigure and manage itself to meet the varying source and load conditions.The research draws inspiration from recent advances in Approximate Dynamic programming to provide real-time and optimal control of the power delivery network under highly dynamic conditions. Hardware based controllers will be developed to provide real-time optimization of the embedded regulators and DC-DC converters for maximum energy-efficiency under performance constraints. The success of this approach is pivoted on advances in the power delivery network, which will also be explored. Traditional ?static? designs cannot be controlled on the fly. Hence, the second principal theme of the research is to explore ?variable structure control? as a means of realizing an ultra-fast and dynamically reconfigurable power delivery system. This will enable orders of magnitude improvement in energy efficiency across wide dynamic ranges of operation and allow new applications for IoT devices with far reaching societal impact.