Specifications
Product type
SOCAMM2 — next-generation server memory module based on LPDDR5X DRAM
Positioning
Brings LPDDR-class power efficiency into a modular, serviceable server form factor
Stacks per module
Up to 4 high-density LPDDR5X stacks per SOCAMM module
Dies per stack
Up to 16 memory dies per LPDDR5X stack
Module capacity
Up to 128 GB in a four-stack configuration, comparable to mid-to-high capacity RDIMMs
Capacity roadmap
Micron has publicly sampled SOCAMM2 modules up to 192 GB
Channel organisation
64-bit channels across the stacked dies
Bandwidth profile
High bandwidth with substantially lower power than DDR5 RDIMM at comparable capacity
Form factor benefit
Compact footprint improves airflow and allows denser memory population around the CPU socket
Thermal
LPDDR5X signalling at lower voltage reduces DIMM-level heat load inside dense AI servers
Target architecture
AI server platforms designed around LPDDR5X host memory such as NVIDIA Grace-based systems
Standards context
SOCAMM is being positioned as a new server memory standard alongside conventional RDIMM and MRDIMM
Ecosystem
Supported by Samsung and Micron with module-level sampling underway for AI infrastructure
Applications fit
AI inference and training servers where memory power and bandwidth efficiency both matter
Overview
SOCAMM2 is Samsung's modular take on LPDDR5X for servers. Standard LPDDR5X is soldered down, which delivers excellent bandwidth per watt but makes capacity fixed at system build. SOCAMM2 packages the same LPDDR5X DRAM into a compact, serviceable module, so AI servers can get LPDDR-class power efficiency without permanently committing the memory configuration.
Each module integrates up to four high-density LPDDR5X stacks with up to 16 dies per stack, organised across 64-bit channels and reaching 128 GB in a four-stack configuration — comparable to a mid-to-high capacity RDIMM. Micron has already sampled modules up to 192 GB, indicating the form factor is not limited to niche low-capacity configurations.
The appeal for AI infrastructure is power and airflow. LPDDR5X operates at lower voltage than DDR5 RDIMM, cutting both module heat and the energy spent moving data between CPU and memory, while the compact footprint lets servers populate more memory around the socket. That combination is what makes SOCAMM2 relevant to inference servers where host memory capacity and power both constrain density.
Key Benefits
LPDDR-class power efficiency in a serviceable module cuts host memory energy and module heat in dense AI servers. Up to 128 GB per module and 192 GB sampling matches RDIMM-class capacity while retaining the power advantage. Compact four-stack construction and 64-bit channel organisation improve airflow and enable higher memory population around the socket — directly useful in Grace-class LPDDR5X server platforms and inference hosts where memory power competes with accelerator budget.
Applications
AI inference and training servers; LPDDR5X-based host platforms such as NVIDIA Grace systems; memory-bandwidth-bound analytics; dense rack deployments where memory power and cooling are constrained; and cloud AI infrastructure looking for capacity-flexible low-power host memory.
Request a Quote — SAMSUNG SOCAMM2 — LPDDR5X SERVER MEMORY MODULE FOR AI DATA CENTERS
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →