High-performance interconnection network is the key in high-performance computing (HPC) and artificial intelligence (AI). Its performance directly affects the efficiency and scalability of the HPC&AI system. With the convergence trend of high performance computing and artificial intelligence computing, the converged interconnection networks for AI&HPC becomes an important trend. With the slowdown of the development of Moore's Law, high-performance interconnection networks face enormous challenges in terms of power consumption, density, scalability, reliability, and cost. It is necessary to adopt new packaging technology and new design technology to design and implement high-performance interconnection network for HPC. The report first analyzes the state-of-the-art representative high-performance interconnection network in the Top500 list, including Nvidia Infiniband, Intel Omni-Path, Cray Slingshot/Aries, and Custom/Proprietary networks including Fugaku Tofu, Bull BXI, TH-Express et al. Based on the analysis of the challenges faced by high-performance interconnection networks design in the post-Moore era and the exascale computing era, this report finally presents the perspective of high-performance interconnection networks.

Lu Pingjing, an associate researcher at the School of Computer Science at the University of National Defense Technology, has been engaged in the development of high-performance computer systems for over ten years. She has participated in the development of multiple generations of Galaxy/Tianhe high-performance computer systems, mainly focusing on computer system architecture and high-performance interconnection networks. She has received two first prizes of the military science and technology progress Award, published over 40 academic papers, more than 40 invention patents, including 4 American patents.