在低電壓環境之物聯網裝置執行二值化神經網路所設計之記憶體架構

Abstract

隨著物聯網的成長，在未來的物聯網裝置端將不會只有做簡單的資料收集以及過濾，也將執行一些簡單的應用，如動作偵測、行為辨識等，此類應用將會以分類為主。而隨著未來傳感器節點端收集的資料的複雜化，其分析難度將會遠超過現在，所以在未來物聯網上將會需要機械學習來協助進行分析。在未來，物聯網裝置也將能利用一些簡單的模組來做辨識，而其學習的部分則由上層的其他裝置進行。但在一般都是執行在低電壓環境下的物聯網裝置，要執行人工神經網路模組這樣的辨識將會有許多額外的消耗，因此我們想要在物聯網裝置上找尋可以降低其執行時之能耗的方法。在分析過成發現其大部分的能耗都來自於低電壓下過長的記憶體存取時間產生的，並非來自於記憶體本身，而是其他設備需等待資料回傳所造成的額外銷耗。因此我們以實際存在的晶片LPC18A1為目標，分別提出了預存取及緩衝控制兩種方式以加速記憶體在低電壓下的存取。雖然記憶體部分的耗能會有些微的增加，但卻能降低整個系統所花的能量，並縮短整體的執行時間。

In the future of IoT, IoT devices will not only perform data collection and filtering, but execute some simple applications, such as motion detection, behavior identification, etc. due to the complexity of the data collected by the sensor nodes in the future. It is too difficult to analyze with human ability. So machine learning will be needed in the future networking to assist in the analysis. Current IoT device is working in the low voltage environment in general; implement the artificial neural network module to such identification will cost additional energy consumption. We attempt to find the methodology that can reduce energy consumption of IoT device in low voltage environment. After analyzed the neural network model (eBNN) on IoT device, we find out most of energy originated from long memory access time in low voltage. However, dominator is not the memory; the dominator is other component that have to wait the data to be read back from memory. We refer the real IoT device chip LPC18A1. In addition, proposed two methodologies, which are buffer control unit and pre-fetch unit to solve this issue. Although the energy consumption of the memory is slightly increased, but the total energy consumption of the IoT device can be reduced along with execution time.