SEOUL, October 13 (AJP) - Samsung Electronics unveiled a joint AI memory expansion solution with U.S. chipmaker AMD to address growing bottlenecks in AI inference, building on a longstanding alliance highlighted by AMD chief executive Lisa Su’s visit to Seoul last week.
At this year’s Open Compute Project (OCP) Global Summit, Samsung demonstrated a solution based on Compute Express Link (CXL), an interconnect technology that allows processors to access additional memory beyond a server’s conventional capacity.
The collaboration targets rising memory requirements as AI agents take on more complex tasks. Inference — the process through which trained AI models generate responses or perform tasks — requires systems to retain and retrieve increasing amounts of information.
Samsung and AMD tested CXL memory as a way to expand capacity and bandwidth, reducing delays caused when processors must wait for data.
Samsung also showcased memory pooling using its CMM-D memory modules, allowing resources to be allocated where needed rather than remaining tied to individual systems. Its MD310 platform enables multiple servers to share a pool of memory, improving utilization as workloads change.
The demonstrations reflect a broader industry shift toward improving how processors, memory and storage work together, alongside advances in individual chip performance.
Samsung organized its OCP exhibition around two areas — “Tokenization Solutions” and “KV-Cache Solutions” — covering memory and storage technologies for AI processing and inference.
The first section featured next-generation high-bandwidth memory, including HBM4E, an HBM5 concept model and custom HBM concepts. Samsung also displayed LPDDR5X memory with processing-in-memory capabilities, which move some computation into the memory itself to reduce data transfers.
Other exhibits included 2-nanometer gate-all-around wafers, Groq language processing unit wafers and Nvidia Groq 3 LPX systems, illustrating partnerships spanning advanced chip manufacturing and AI computing.
The KV-cache section focused on managing information that AI models retain during inference. A key-value cache stores results from earlier processing steps so a model can reuse them instead of repeating calculations.
Samsung presented CXL memory, enterprise solid-state drives and software designed to store and retrieve that data more efficiently. Demonstrations included memory pooling based on MD310, AI Memory Node, vector database and direct-storage technologies, alongside liquid cooling.
The company also outlined plans to expand joint development through its Memory Application eXperience Lab, or MAX Lab, where customers and partners can evaluate emerging memory and storage technologies under real-world workloads.
The lab supports testing and system design involving processing-in-memory, CXL memory expansion and next-generation solid-state drives, allowing customers to assess technologies before deployment and incorporate their requirements earlier in development.
Samsung is also expanding open-source software development and standardization through its Global Open-ecoSystem Team, or GOST, with work on KV-cache optimization and direct storage.
Together, the initiatives point to a broader role for Samsung in AI infrastructure: supplying memory chips while helping customers integrate them into systems capable of handling increasingly demanding inference workloads.
Copyright ⓒ Aju Press All rights reserved.
