ETRI Expands Computing Resources with Light Opening a New Era for AI Datacenters

A team of Korean researchers has developed the world’s first technology that can freely connect and disconnect core computing resources such as memory and accelerators with “light” in next-generation artificial intelligence (AI) datacenters.

Electronics and Telecommunications Research Institute (ETRI) announced the development of a new optical switch based datacenter resource interconnection technology (Optical Disaggregation, OD).

This technology is regarded as a core next-generation optical network technology that is designed to resolve the shortage of computing resources due to the increasing AI services and that enables faster and more efficient operation of future datacenters.

Current datacenters are designed based on a server-centric architecture in which CPUs, memory, storage, and accelerators (GPUs) are bundled within a single server.

In this architecture, only the limited resources installed in the same server can be utilized, leading to problems in which overall efficiency drops due to large deviations in resource utilization, such as some servers using only memory excessively while others use only the CPU.

In addition, most datacenters use electrical packet-based switches, which causes delays as multiple optical-electrical-optical conversions occur during the data switching process. These delays limit the performance of delay-sensitive operation such as connecting remote memory resource and large-scale AI training.

The Optical Disaggregation (OD) technology developed by ETRI fundamentally solved these structural limitations. When memory or accelerators within a server are insufficient, an optical switch can be used to instantly connect remote resources to the server via an optical switch.

This enables resources to be connected and disconnected rapidly and flexibly, exactly when needed and in the exact amount required, even for tasks demanding high-performance computing such as AI training or large-scale data analysis.

In particular, this technology is significant in terms of global technological competitiveness as it is the world’s first case of connecting the Compute Express Link (CXL) standard with an optical switch.

The research team successfully demonstrated the technology by building a verification system combining ETRI’s in-house developed CPU adapter, memory blade, accelerator blade, and OD manager.

The experiment confirmed that when a service program requests additional resources, the optical path is automatically set to allocate the necessary memory and accelerators in real-time, and the service is steadily provided. In other words, they demonstrated for the first time in the world a system that can provide light speed connection between datacenter resources using software control.

ETRI secured original patents related to CXL applied to this technology and filed 47 domestic and international patents for related technologies. In addition, the research results were presented at the Optical Fiber Communications Conference and Exposition (OFC) and the European Conference on Optical Communication (ECOC), the most prestigious conferences in the field of optical communication, gaining recognition for its technological prowess in the international academic community.

Furthermore, this research officially proved its excellence by being included in the “Top 100 Excellent National R&D Performances” selected by the Ministry of Science and ICT in 2023.

Lee Jun Ki, Director of Optical Network Research Section at ETRI, said,

“Datacenter resources are being consumed rapidly worldwide due to the increase of AI services. This research achievement will serve as an important opportunity to resolve the datacenter resource shortage problem by efficiently sharing and utilizing memory and accelerators, and to accelerate the transition to sustainable future datacenters,”

ETRI plans to apply this technology across a range of fields, including the advancement of national AI infrastructure, integration of cloud and supercomputing, and the development of eco-friendly datacenters for driving innovation across industries.