Fast Optical Switching for Computing-Power Interconnects

With the development of Artificial Intelligence(AI)technologies such as large language models and the promotion of major national projects such as "Naional Somputing Network to Synergize East and West", the internal network interconnection of Artificial Intelligence Data Centers (AIDC) and Computing-Power Networks (CPNs) has become an important tendency of optical network development. Compared with the traditional data-oriented interconnection, the AIDC networks and CPNs have a higher demand for transmission delay, throughput and determinism. On the one hand, high density, low-latency processing and computing resource scheduling are required within the AIDC to meet large-scale real-time training and reasoning. On the other hand, interconnection between the AIDCs requires high-bandwidth, low-latency, high-reliability data transmission and resource sharing to support multi-data-center collaboration and failure recovery. Therefore, the Computing-Power Interconnection (CPI) needs optical switching technology to support efficient data transmission in the optical transport network within the AIDCs and fast bandwidth scheduling between the AIDCs. Hence, constructing a CPI-oriental optical switching network can accommodate the AIDC network requirements and provide efficient, deterministic, and reliable data services. In this paper, we analyze the optical switching technology for AIDC networks, including CPI architecture, data switching and application scenarios.