AI Technology in Networks-on-Chip

Abstract

The on-chip network, commonly known as network-on-chip (NoC) on a die as an alternate prevalent interconnection infrastructure, has been continuously occupying the space of systems-on-chip (SoCs). In other words, the SoC-based systems are continually transformed to NoC-based multiprocessor and multicore designs. These designs would prospectively facilitate high core utilization and the need for high performance. To date, performance evaluation of NoCs is largely based on simulation employing a set of NoC simulators, e.g., Noxim, BookSim, etc. However, the use of simulators has several limitations. Therefore, it has become essential to opt for an alternate approach. The rapid development of the artificial intelligence (AI) industry currently leads the next generation of computation. Consequently, AI technology in every aspect becomes an essential milestone in surpassing the human level. Recent research shows that AI technology, including machine learning, deep learning, etc., is often preferred to improve data transformation and hardware efficiency in on-chip communication networks. Such networks are called AI chips, where an NoC is one feasible solution to SoCs. This chapter seeks to critically situate the AI technology in evaluating the NoC performance at different perspectives: theoretical, technological, analytical, etc. The chapter elucidates some of the salient problems, interests, and issues for its organization around topology exploration and design considerations in NoCs; customization of NoCs for AI chips; performance evaluation via simulation; performance predictions using machine learning, deep learning, etc. The overall thrust of this chapter is to provide “Core Perspective” and “Further Perspective” of the application of AI technology in predicting several NoC performance metrics toward high-speed communication and voluminous computation by multiprocessor systems-on-chip. Thus, one can see how this AI technology dramatically improves the speed and accuracy of predicting various performance metrics.