2013-08-012024-05-18https://scholars.lib.ntu.edu.tw/handle/123456789/706384摘要：隨著近年來雲端運算技術以及應用的發展，未來大部分的電子裝置都將隨時與網路連 結，運算以及資料都將分散在雲伺服器(cloud server)和終端裝置(terminal device) 上，如此的運算架構被稱為連結型的運算架構(connected computing)。在如此的架構 之下，一些需要大量運算以及大量資料的工作，將會由運算能力極強的雲伺服器來負 責，而行動終端將會保留介面的部分，而人機介面也將從單純的輸入輸出轉變為更具 智慧功能、更著重於視覺感受的智慧型人機介面(smart interface)，本計畫的主題， 便是期望能夠模仿人類新皮質的運作方法，設計出智慧型人機介面的處理晶片以及相 關處理元件架構以及演算法設計。基於我們對於人類新皮質運算方法的了解，以及以 往在可重組化影像串流處理器、繪圖晶片、辨識晶片設計的經驗，在本計畫之中，我 們將會發展一智慧型人機介面之處理晶片及其相關之模組架構以及演算法設計，此晶 片預計將由處理器陣列、串流處理器陣列、機器學習加速器、注意力模型運算、運算 記憶體等元件組成，而這些運算單元間的連結以及運算單元均為可重組化，成為一異 質性微型雲量級運算架構 (heterogeneous micro-cloud computing architecture)， 運算單元間的溝通以及資料的傳輸，將可能使用可重組化串流處理(reconfigurable stream processing)的方式或是以排程器(scheduler)統一控制的方法。<br> Abstract: With the advances of cloud computing technique and applications, in the future, most of the devices will connect with Internet, where the computation and data will be distributed between cloud servers and terminal devices. In this connected computing architecture, tasks requiring large amount of computation and data will be located in the powerful cloud servers, while the terminal devices will take charge of the human-machine-interface (HCI). “Smart interface,” where intelligent visual interaction is employed instead of keyboard and mouse, will become more and more popular. In this project, inspired by the neocortex computing framework of humans, we plan to develop a new processor and related techniques for smart interface. Based on our understanding about neocortex computing and our experiences in designing reconfigurable image stream processors, graphics processing units, and recognition processors, this chip will be composed of a processor array, a stream processor array, a machine learning engine, an attention model engine, computing memory, and several other accelerators. These various modules will be connected with a reconfigurable on-chip micro network as a heterogeneous micro-cloud computing architecture. Reconfigurable stream processing or centralized scheduler is planned to be employed for inter-module communication and data transmission. This processor is targeted to have the following features: 1) neocortex-inspired computing architecture supporting neocortex-inspired intelligent algorithms for smart interface; 2) heterogeneous micro-cloud computing architecture; 3) control scheme with reconfigurable stream processing or centralized scheduler; 4) having processor array, stream processor array, machine learning engine, attention model engine, and computing memory; 5) interconnection with network-on-chip; 6) supporting connected computing; 7) supporting prediction/synthesis-based recognition; 8) supporting several functions, such as object segmentation/recognition/detection and augmented reality.智慧型人機介面新皮質雲量級運算擴增實境機器學習注意力引擎運算記憶體Smart interfaceneocortexsingle-chip cloud computingaugmented realitymachine learningattention enginecomputing memory