The rapid advancement of Artificial Intelligence (AI) and Machine Learning (ML) technologies has led to a surge in demand for High Performance Computing (HPC) and extensive data processing capabilities. GPU interconnect ports have rapidly upgraded to 800G (2x400G) and are expected to reach 1.6T (2x800G) in the next generation. This significant evolution highlights the increasingly urgent need for higher bandwidth in optical interconnects within data centers.

AI/ML Cluster Interconnection, new Challenges for HPC and Data Center Interconnects

With the rapid growth of AI/ML clusters, expanding from a few hundred to tens of thousands of accelerator cards, the interconnect distance has also increased from tens of meters to over a hundred meters. This has created a pressing demand for low power consumption and low latency, while traditional electronic switching faces bottlenecks. Simultaneously, HPC is confronted with higher requirements, including flexible configuration of cluster size and long-term stable operation of networks. Large-scale data center interconnects also encounter challenges in Spine switch performance bottlenecks and the evolution of port rates. As rates increase, relying solely on hardware upgrades for energy efficiency improvement above 200G may not be significant, necessitating further enhancement through architectural reconstruction to reduce operational costs.

Supporting Software-Defined Networking (SDN), the Emergence of Full Optical Switching Solution

To tackle these challenges, the Full Optical Switching (OCS) solution has emerged, capitalizing on its advantages in complete optical switching. OCS enables transparent transmission of optical signals and supports the exchange of optical signals at any rate, modulation format, or communication wavelength in optical fibers. It boasts features such as zero clock jitter, no delay, no data reading, and no leakage risk. Furthermore, OCS provides support for Software-Defined Networking (SDN), enhancing control and increasing system flexibility.

The successful application of the OCS full optical switching solution is evident in leading data centers. Just on December 6th of last year, with the launch of Google's AI model Gemini and its corresponding TPUv5 hardware version, the utilization of OCS optical switching technology was reinforced. In the case of Google TPUv4 application, OCS notably enhanced the reliability of cluster interconnects. Even with host reliability dropping to 99.0%, it ensured exceptional performance for TPU slices, resulting in an overall system performance improvement exceeding 6 times.

Extensive Applications of the OCS Solution

The Software-Defined Full Optical Switching, renowned for its exceptional performance in high-performance computing and data center applications, plays a pivotal role in supporting the advancement of AI networks. As data centers undergo transformation, the expansive potential applications of OCS full optical switching technology in optical interconnects will invigorate the success of digital transformation. The era of large-scale AI models is poised to instigate a computational arms race, propelling advancements in computational chips and networks and potentially fostering sustained industry growth.

The demand for bandwidth in optical interconnects within intelligent computing data centers is on the rise. With the expansion of cluster sizes, there is a need for larger-port OCS, coupled with enhancements in OCS reliability and insertion loss performance. The OCS solution, utilizing DirectLight direct beam deflection technology, excels in various aspects, including ultra-low loss, high seismic resistance, and substantial benefits in operational costs, throughput, power consumption, and overall expenses. DirectLight-based OCS solutions facilitate dependable remote selective routing, real-time performance monitoring, and automatic protection. Beyond applications in intelligent computing centers, these solutions find widespread use in quantum internet, and other domains, positioning OCS as a critical element in integrated air-space-ground networks during the 6G era.

LUSTER Large Matrix Full Optical Switching Solution

Since 2001, LUSTER has been dedicated to advancing optical switching technology, products, and applications. In 2015, we forged a strategic partnership with the global optical switching leader H+S Polatis. H+S Polatis specializes in delivering low-loss full optical switching solutions for tasks like remote fiber layer configuration, protection, monitoring, reconfiguration, and testing.

The H+S Polatis 6000/7000 series full optical switches showcase exceptional performance, supporting fully non-blocking matrix scales ranging from 8x8 to 576x576, with outstanding seismic resistance suitable for some applications. Boasting ultra-low insertion loss, exceptional optical metrics, support for single-channel bidirectional transmission, and dark fiber switching, these switches provide completely transparent optical transmission. They are protocol-independent, allowing seamless upgrades to rates of 400G/800G and beyond. The optional N×N symmetric and arbitrary port configuration structures (CC series) cater to SDN applications. These switches find extensive application in various scenarios, including AI computing centers with full optical switching, quantum internet, and automated testing in production lines.