Samsung, AMD showcase CXL memory sharing for AI servers at OCP Summit

Samsung Electronics and AMD jointly demonstrated a technology that lets multiple servers share a common pool of memory to meet the growing memory demands of AI inference. 

The technology is Compute Express Link, or CXL. Normally, a server can use only the memory installed inside it, so one machine can run short while another has memory sitting idle. CXL connects servers to a shared pool of memory so that a server running a heavy AI job can borrow extra capacity from the pool and return it when the job is done.

The demonstration took place at the 2026 OCP Global Summit, where Samsung took part as a Diamond sponsor, the top tier. The annual event is run by the Open Compute Project, an industry group where technology, semiconductor, server and data center companies share open hardware designs and discuss standards for AI infrastructure.

Samsung said the demonstration is significant because the technology was validated in an actual system with a global computing platform company, and it expands cooperation on improving AI infrastructure performance.

The company also outlined other directions for raising AI infrastructure performance. They include high-performance memories such as HBM5 and custom HBM, and data storage using next-generation solid-state drives (SSD). HBM, or high bandwidth memory, stacks memory chips vertically so data moves faster.

Shim Ho-jun, a vice president on Samsung’s application engineering team, said at a summit session that memory speed alone is no longer enough as AI systems grow more complex. He proposed three directions: speed, memory expansion and data persistence, meaning the efficient storage and management of large volumes of AI data.

Samsung also ran two exhibition zones. The Tokenization Solutions zone showed seventh- and eight-generation HBM4E, HBM5 and custom HBM concept mock-ups, along with LPDDR5X-PIM, a low-power memory chip that performs some computation itself. This reduces the data moved between memory and processor and improves power efficiency.

The KV-Cache Solutions zone centered on CXL memory, enterprise SSDs, and technologies for storing KV cache, the temporary data an AI model keeps while generating a response.

The summit runs through Thursday in San Jose, California.

BY LEE JAE-LIM [[email protected]]