Innovative Computing for Solving Social Issues

Since the development of practical stored-program computers in the late 1940s, performance has risen amazingly by about 10(12) times over a period of 70 years. However, it is generally recognized that semiconductor transistor scaling is reaching its limits and that Moore's law is coming to an end. Regardless of these technical issues, the explosive increase in the amount of data generated in today's loT era is expected to continue, and it is highly anticipated that this data will be used to create new value and novel services. Meeting these expectations will therefore require improvements in performance independent of Moore's law. To address these issues, Fujitsu Laboratories proposes domain-specific computing as a new computing paradigm. The aim of domain-specific computing is to break through Moore's law by adopting 'architecture specific to the type of processing needed in fields such as knowledge processing whose objective is not to obtain rigorous numerical results. For example, in application to deep learning engines, high-speed image search engines, and machines dedicated to combinatorial optimization problems, domain-specific computing has demonstrated that it showed 50-12,000 times higher performance than that of conventional approaches. In this paper, we describe the direction of domain-specific computing as a new computing paradigm and present specific application examples.