Abstract
摘要:物聯網帶動了許多革命性的轉變與進步,感測資料經由資訊處理所產生的附加價值協助人們做出決策,進而逐步改變人們的生活模式。為了降低物聯網中資料傳輸流量、提升後端資訊分析運算的效率,思科(Cisco)提出階層性「霧運算(Fog Computing)」的概念。但是相對於雲端運算,在階層運算結構下,在接近感知層的節點上所擁有的記憶體容量、硬體運算資源皆非常有限。另一方面,晶片系統內的功率密度在先進製程下呈現指數性成長,但是,電池所能提供的能源卻成長有限,使得功率缺口越來越大,因此能源效率將會是首要議題。
為解決上述問題,本計畫致力於智慧型綠能物聯網系統關鍵技術開發(Enabling Technologies for CErebral and GREen IoT System, CERES),將著重於兩個主軸方向來發展:1)物聯感測資料擷取層面的高能源效率智慧感測器電路模組設計,以及 2)物聯感測資訊分析層面的微型綠能智慧機器學習系統開發。
本計畫提出六項主題進行研究,包含1)智慧節能物聯感測訊號處理晶片,探討近臨界電壓(Near-Threshold Voltage, NTV)電路設計,加入動態調整及可重組化運算設計方式,讓高能量使用效率之操作點能廣泛應用於各種資料擷取設備。2)低功耗時脈產生器模組開發,探討石英振盪器與鎖相迴路此兩大重點。3)超低電壓電路之進階測試技巧,提出針對動態電壓和頻率縮放(DVFS)低功耗設計之先進多模式測試產生技巧,以及使用機器學習針對極低功率電路之進階壓降預測技巧。4)新式光學感測器資料處理,針對動態視覺與壓縮感測此兩大類新式光學感測器資料處理技術做開發,並將處理演算法以硬體進行計算加速。5)強健(Robust)且輕量(Light)機器學習引擎,開發在壓縮後的資料中進行學習之技術,並利用機器學習演算法中固有的容錯性質,提出適應性錯誤感知計算模型來達到強健式學習的目標,最終落實於資源有限之運算層。6)高效率熱電轉換技術,開發高熱電轉換效率的材料,用以回收運算平台散失的熱能,藉此彌補運算節點有限資源之困境,來達到綠運算平台之目標。以上針對高能源效率設計的進階感測模組與微型綠能機器學習系統開發等六項子計畫,可以分別解決未來智慧型綠能物聯網系統基礎技術開發之挑戰。
Abstract: The Internet of Things (IoT) speeds up awareness and response to events. It is transforming whole industries by helping us to make decisions and changing our life style. However, the cloud by itself cannot connect and analyze data from thousands and millions of different kinds of things. Therefore, Cisco proposed Fog Computing to reduce the amount of data transmission and increase the computation efficiency of data analysis. However, as the amount of layers increases, the memory and computation hardware resources on fog nodes become limited compared to the cloud. Additionally, the growth of the power density on SoC is exponential in advanced process, but the growth of the power limit produced by batteries is much slower. This causes the power gap to become larger and larger. Hence the energy efficiency problem will be the prior issue.
To solve the above problems, this project is committed to develop “Enabling Technologies for CErebral and GREen IoT System, CERES.” Our team will focus on two main directions: 1) The design of high energy efficient and smart sensor hardware modules, and 2) The development of the green and light-machine learning system on data analysis layers. This group project contains six topics, including
1) Machine-Type-Oriented Digital Circuit Design: probing to design near-threshold voltage hardware with dynamic and reconfigurable adjustment, letting the high-energy-efficient operating point be suitable for different kinds of sensors.
2) Low-Power Clock Generator Design: containing Crystal Oscillator (XO) and Phase-Locked Loop (PLL).
3) Advanced Test Techniques for Intelligent Energy Efficient Circuits: proposing the method of low-power multi-mode test generation against DVFS and the method of voltage-drop prediction to ultra-low power hardware by machine learning.
4) Algorithms for Novel Optical Sensor Data Processing: developing data processing algorithms for new optical sensors based on Dynamic Vision and Compressive Sensing technologies. Implementing hardware accelerators for the proposed data processing algorithms.
5) Light and Robust Machine Learning Engine: developing techniques of learning from compressed data and error-aware computation model with error-resilient properties of machine learning to achieve robust learning. This system will be implemented in the resource-constraint fog nodes.
6) Thermoelectric Energy Harvesting: developing materials with high thermoelectric transformation efficiency to recycle thermal energy lost from the computation platform and help the resource-constraint fog nodes.
These six sub-projects focus on developing advanced sensors with high energy efficiency and the green light-machine learning system, solving the future challenges of fundamental techniques development for the intelligent and green IoT system.
Keyword(s)
物聯網
低功耗電路設計
進階電路測試技術
機器學習
Internet of Things
Low Power Circuit Design
Advanced Test Techniques
Machine Learning
Thermoelectric Energy Harvesting