Micron begins sampling 192GB SOCAMM2 low-power memory for AI servers

Micron shows SOCAMM2 memory, 192GB per module

Micron has announced sampling of its new SOCAMM2 memory modules, offering up to 192GB capacity per module based on LPDDR5X. These modules are designed for data center systems focused on AI workloads, where power efficiency and memory bandwidth are becoming core requirements. SOCAMM2 is an updated version of Micron’s earlier SOCAMM design, now offering 50% more capacity within the same footprint.

Source: Micron

The company states that increasing memory capacity in this way can reduce time-to-first-token in real-time inference tasks by more than 80%. SOCAMM2 is manufactured on Micron’s 1-gamma DRAM process, which targets over 20% lower power usage than the previous generation. Large AI server racks equipped with multiple terabytes of low-power memory could see considerable total energy savings.

Source: Micron

Micron worked with NVIDIA on incorporating low-power DRAM into server designs, and SOCAMM2 is intended for platforms handling large context window AI models. The modules use LPDDR5X to deliver high bandwidth while maintaining reduced power draw, which helps manage system-level thermals.

Micron claims SOCAMM2 can achieve more than two-thirds lower power usage compared to equivalent RDIMM configurations, while occupying about one-third of the board area. The compact size may also benefit liquid-cooled server layouts.

SOCAMM2 is currently sampling to customers at up to 192GB per module and speeds up to 9.6 Gbps. Mass production will begin according to partner launch timelines. Micron did not indicate any plans for consumer or desktop use, and the technology remains targeted at data center platforms only.

Source: Micron PR