Samsung Electronics and AMD jointly demonstrated a Compute Express Link memory expansion system at the OCP Global Summit in San Jose, the companies said.
The demonstration used memory pooling, a technique that lets several systems share CMM-D modules, the CXL memory product Samsung has been developing, according to workload. Operators gain capacity without buying more server nodes. The two firms first signed a memorandum of understanding on next-generation AI memory cooperation in March.
Samsung executive vice president Shim Ho-joon framed the strategy around three priorities in a keynote: speed, memory expansion and data persistence.
The pitch reflects a shift in the bottleneck. As AI agents take on multi-step tasks, the key-value cache that stores earlier context has grown, making capacity as important as throughput for inference servers. Samsung said its exhibit showed HBM4E and HBM5, custom HBM concepts and low-power DDR5X with processing in memory, alongside 2-nanometre gate-all-around wafers produced by its foundry.
SK hynix presented a memory-centric rack design at the same event. Samsung expects to bring the CXL capability to customers through its MAX Lab, where pre-commercial parts are tested under production workloads.
