Implementation of the CXL Ultra-Large Memory System through Trilateral Cooperation

As demand for AI, high-performance computing, and big data analysis increases, the focus of server performance competition is shifting from computing units to memory structures. To respond to this change, Samsung Electronics, the Electronics and Telecommunications Research Institute (ETRI), and PrimeMars have embarked on the joint development of a CXL-based ultra-large capacity memory solution and are working to implement a memory-centric computing structure into actual systems.

Samsung Electronics, ETRI, and PrimeMars officially announced their collaboration in Seoul on April 6 to develop a massive memory system utilizing CXL (Compute Express Link), a next-generation interconnect technology. This collaboration focuses on verifying a structure that flexibly expands and shares memory resources within data center infrastructure, going beyond simply increasing memory capacity.

Existing servers have been designed around computing units such as CPUs, GPUs, and TPUs, but as data-intensive tasks increase, they are revealing limitations in memory scalability and resource utilization efficiency. CXL is a technology that connects memory in a fabric format to enable wider utilization of necessary resources, and it is discussed as an alternative to reduce such bottlenecks.

In this development, Samsung Electronics provides high-capacity DRAM, the latest DDR DIMM technology, and memory fabric management and orchestration software capabilities. PrimeMars is responsible for a memory expansion solution in the form of an Add-in Card (AIC) utilizing the chiplet-based CXL controller SoC 'Falcon'. ETRI is responsible for combining the technologies from both sides to implement a complete system in the form of an ultra-large capacity memory chassis.

The cooperation among the three organizations is an extension of existing bilateral collaboration. Samsung Electronics and PrimeMars signed a memorandum of understanding in August 2024 for the joint research and development of advanced CXL solutions, and with the addition of ETRI, the scope has expanded to the system integration stage. Accordingly, the initiative has taken on the nature of testing whether CXL memory technology can be connected to actual infrastructure structures beyond components and modules.

The three companies plan to jointly pursue the demonstration of large-scale memory pooling technology, the advancement of memory management technology, and the implementation of a memory-centric architecture capable of handling AI workloads. The industry is watching closely to see if this collaboration will serve as an opportunity to enhance the technological self-reliance of the domestic CXL ecosystem and expand the role of Korean companies and research institutions in the transition to AI data centers.