Emerging Hardware Technologies for IoT Data Processing

Abstract

Fast and energy-efficient data processing has become a critical need for various forms of computing in the era of Internet of things (IoT). Emerging IoT applications demand for increasingly high data collection rates and significant computational requirements that often do not fit in the stringent power envelopes of the existing IoT devices. Data centers and cloud servers are then used to empower the IoT systems by performing massive data processing on behalf of the IoT users. Recent years have witnessed many significant challenges for big data processing in IoT systems. This section provides an overview of main architectural challenges for data processing in IoT systems and explains a number of recent innovations for addressing the challenges. The rest of the section examines two example memory architectures for accelerating data-intensive applications in the IoT nodes and data centers. The first architecture is a memory subsystem based on the emerging nonvolatile memory technologies that enables energy-efficient neural network acceleration in memory arrays. The memory system is capable of performing ordinary data storage in the future IoT nodes, as well as significantly accelerating certain operations for binary neural network workloads. The second architecture is a memory-centric accelerator specifically designed to perform large-scale data clustering using the k-median algorithm. The accelerator is suitable for IoT data centers, where large data is collected from the IoT nodes and clustered in the cloud servers. The architecture has shown significant energy-efficiency and performance potentials for gene expression analysis from the healthcare sector and document clustering used for data mining in web applications.