–
Room: TBD
The rapid evolution of artificial intelligence (AI) technologies with numerous parameters, including large language model (LLM)-based services, presents critical issues on yet further scaling the network and interconnecting data centers in the age of AI. Recently, co-packaged optics (CPO) and optical switching have attracted considerable attention to address the bottlenecks of I/O and switching in data center networks. Consequently, a question has arisen whether advances in CPO, including optical-layer switching, will enable further scaling of AI clusters.
Another important issue pertains to the operations of various LLM-based services that process enormous amounts of data. Most of these data are collected through Internet-of-Things (IoT) devices or mobile networks, while cloud data centers tend to be situated in locations remote from edges. Indeed, with a continuous increase in the volume of edge data, moving and storing data not only consumes significant amounts of energy but also incurs undesirable latency levels. This will lead to an issue in regard to how and where such data are collected, transferred, stored, and processed for training and inference tasks. Then, the compute and network resources must not be treated separately but optimized holistically by redressing the allocation of edge and cloud computing. To achieve this, the flow of data must be optimized by identifying the roles of CPO and optical-layer switching at all levels of interconnect, from the chip-scale, on-board, rack-scale, and intra-/inter-data center (DC) levels to even mobile-edge and wide-area levels. Thus, such an approach may completely change the landscape of digital infrastructure.
This workshop will bring together AI-DC architects and optical networking experts from industry and academia to discuss the future of AI-related digital infrastructure, expectations, and target specs of emerging optical network technologies, including CPO and optical-layer switching. The workshop comprises two technical sessions followed by a panel. In the first session, requirements for networks and interconnects will be discussed from the perspectives of various systems, whereas the second session will discuss CPO and optical switch technologies. The panel will serve as a cross-examination between the two sessions.
Organizers
Nicola Calabretta, Technische University, Netherlands
Shu Namili, AIST, Japan
Mahdi Nikdast, Colorado State University, United States
Speakers: Session 1
Keren Bergman, Columbia University, United States
Hong Liu, Google, United States
Shintaro Mizuno, NTT, Japan
M. Ashkan Seyedi, NVIDIA Corporation, United States
Speakers: Session 2
Joris Van Campenhout, IMEC, Germany
Katsumi Fukumitsu, Fujitsu, Japan
Ling Liao, Intel Capital, United States
Peter Winzer, Nubis Communications, United States
Ming Wu, University of Berkeley, United States