Micron Unveils 256GB LPDDR5x Module for AI Servers

Micron has introduced a 256GB SOCAMM2 memory module based on LPDDR5x chips, explicitly aimed at AI server and hyperscaler deployments. The module is built using 64 stacked 32GB LPDDR5x dies, packing a large capacity into a compact form factor designed for dense memory configurations.

One notable point is how hyperscalers can use these modules: Micron says up to eight of the 256GB units can be installed in a single AI server to reach 2TB of system memory. That kind of density appeals to large-scale inference and training workloads that benefit from keeping more data and model parameters in fast memory close to accelerators.

Beyond capacity, Micron highlights power efficiency improvements. LPDDR5x is already tuned for lower power per bit compared with older DRAM generations, and the SOCAMM2 design aims to reduce overall server power draw — a key concern for data centers running energy-hungry AI jobs.

These modules aren’t aimed at everyday PCs or mainstream servers. Instead, they’re targeted at hyperscalers and enterprises that operate specialized AI hardware and require both high capacity and efficiency. For most organizations and individual buyers, the scale and cost of such modules will likely be out of reach.

From an operational perspective, the introduction of high-capacity LPDDR5x modules further narrows the gap between accelerator-local memory needs and what system memory can provide. For teams planning large-model deployments, this could simplify system design by offering dense, energy-efficient RAM options without relying solely on expensive on-package memory.