Comparative study of the domain-specific processors

Abstract

Microprocessor is the core component in modern information systems. The explosive growth of data volume, fueled by the rapidly evolved technology including big data, artificial intelligence and 5G, demands an unprecedented high computing capability. It has been believed that domain-specific computing will be one of the major directions of computer architecture in the post-Moore's Law era. The evolution of domain-specific processors is actually always accompanied by general-purpose processors. Digital signal processor (DSP), one of the typical domain-specific architectures, appeared even earlier than common CPU technology. General-purpose computing, which has been witnessed to achieve huge commercial success, also shares lots of key technologies to boost domain-specific architectures, in terms of performance and programmability. By surveying representative commercial products of DSP, GPU, AI chips and NPU, this paper aims to explore the key architectural features of these domain-specific architectures, and is supposed to shed some light on the future research and development of domain-specific computing technology.