GreenArmy：綠色微雲伺服系統晶片平台技術-子計畫六：綠色微雲伺服系統晶片上全數位功率及熱能偵測與管理之硬體設計( I )

Abstract

本計畫欲支援多核心系統之偵測與控制管理硬體架構，此子計畫的主軸在於結合嵌入式多核心系統設計實現軟體定義多輸出入鎖相迴路(software-defined multi-input multi-output phase-locked loop； MIMO SDPLL) 設計，並實現可支援 Cloudlet server on chip 之相關偵測與控制管理所需硬體，例如：溫度偵測、電流偵測、快速時脈調整及電壓開關，以提供不同virtual machine (VM)對性能需求與省電目標。本子計畫在技術上主要分為三個方面研究議題著手，其中包含全數位溫度偵測器(all-digital thermal sensor)、全數位電流偵測器(all-digital current sensor)與軟體定義多輸出入鎖相技術(MIMO-PLL)。此架構特色在 (1) 軟體定義多輸出入鎖相迴路(MIMO SDPLL): 將多個全數位偵測器(TED)和全數位震盪器(DCO)結合中央處理器以形成軟體定義多輸出入鎖相迴路軟體，此鎖相迴路不需使用其他運算單元，以降低實際PLL 硬體的套數。 (2) Shared All-Digital Thermal and Current Sensor: 開發可同時支援SDPLL 所需時脈輸出、相位偵測和時間偵測，並具備溫度和電流偵測功能之矽智產(silicon IP)以提供管理軟體使用。 (3)利用分享Cloudlet server on chip 的處理核心同時處理多個MIMO-PLL 運算，並分配其他不同核心所需之不同時脈的控制，藉由管理軟體的動態分配MIMO-PLL 運算可進一步提高多時脈輸出入的運作效率。

This project mainly focuses on software-defined multi-input multi-output phase-locked loop (MIMO SDPLL) to support all-digital sensing and multi-rate clocking techniques and hardware implementations for GreenArmy-Cloudlet server on chip, such as thermal detection, current detection and fast-adjust clocking, to ensure different virtual machines (VM) balancing power and performance. There are three research topics to proceed, which are all-digital thermal sensor, all-digital current sensor and fast adjustable multi-rate clocks. Our features are (1) MIMO SDPLL: integrate several timing error detectors (TED) and the all-digital oscillators (DCO) with the embedded processor to support multi-clock operations. By using software flexibility, this kind of PLL methodology can reduce the actual copies of PLLs in multi-rate clocking. (2) Shared all-digital thermal and current sensor: achieve low complexity to make TED and DCO sense thermal and current. Then an MIMO SDPLL can sense thermal and current without hardware overhead. (3) Dynamic allocation of MIMO-PLL processing in Cloudlet server on chip: use the software management to locate different MIMO-PLL processing in different cores dynamically to enhance core utilization and to improve power efficiency.