Samsung explains SOCAMM2: LPDDRX memory for AI servers

Samsung LPDDR5X in SOCAMM2 footprint

Samsung has announced SOCAMM2, short for Small Outline Compression Attached Memory Module. It is a server memory module that uses LPDDR, a low-power memory type usually found in laptops and phones.

SOCAMM2 is meant for AI data center servers, where systems run training and constant inference workloads and need fast response while keeping power use down. Samsung says it is already supplying samples to customers.

Traditional servers use DDR memory on RDIMM modules. SOCAMM2 takes LPDDR and puts it into a detachable, modular server format. Samsung claims SOCAMM2 delivers more than 2x the bandwidth of RDIMMs and cuts power use by 55%, while keeping performance stable under heavy AI load.

Source: Samsung

SOCAMM2 is designed to fix a common server LPDDR drawback. LPDDR in servers has often been soldered to the motherboard , which makes upgrades and replacements harder. With SOCAMM2, the memory can be swapped like a normal module , which should reduce downtime and lower maintenance costs.

Samsung also says SOCAMM2’s layout is horizontal rather than the usual vertical DIMM placement. The goal is better space use and more flexibility for heatsinks and airflow, with support for air and water cooling designs.

NVIDIA and JEDEC standardization

Samsung says it is working with NVIDIA to optimize SOCAMM2 for NVIDIA accelerated infrastructure, aimed at inference platforms. Samsung also says JEDEC standardization is close to completion, which should help with broader platform support over time.

SOCAMM2 is not aimed at consumer PCs. It is a server-only memory module for AI infrastructure.

Source: Samsung