Samsung unveils SOCAMM2 low-power memory for AI data centers

Samsung has introduced SOCAMM2, a new LPDDR-based memory module for AI data centers, with customer samples already being supplied.

SOCAMM2 is designed to support AI server workloads that require high bandwidth and power efficiency, combining LPDDR5X DRAM with a modular, detachable form factor.

Samsung said the module offers more than double the bandwidth of traditional RDIMM server memory, and uses over 55% less power.

The detachable design allows for easier upgrades and maintenance, potentially reducing downtime and total cost of ownership for operators.

Samsung is collaborating with Nvidia to optimize SOCAMM2 for Nvidia’s accelerated infrastructure.

Efforts are also underway to standardize LPDDR-based server modules in the industry, with Samsung contributing to these initiatives.

SOCAMM2 is aimed at addressing growing demand for energy-efficient, high-performance memory as AI workloads increase in data centers.

🔗 Source: Samsung

🧠 Food for thought

Implications, context, and why it matters.

SOCAMM2 speed depends on signal integrity and build quality

JEDEC, the memory-interface standards body, expects SOCAMM2 to support LPDDR5X at up to 9.6 Gb/s per pin ‘where signal integrity allows’ 1. Real speed hinges on printed circuit board (PCB) layout or thermal management, so AI servers may miss peak rates.
Samsung’s ‘more than double the bandwidth’ claim likely assumes optimal conditions. Nvidia reportedly dropped first-generation SOCAMM modules after overheating on its Blackwell Ultra GB300 ‘Cordelia’ platform (an AI accelerator platform) 2, which exposes thermal and signal integrity hurdles in compact memory.
The article emphasizes work with Nvidia to avoid issues. Deployment depends on whether server original equipment manufacturers (OEMs) can run 9.6 Gb/s reliably in production chassis, which could cap near-term gains.

SOCAMM2 connector and thermal suppliers see openings

SOCAMM2 uses Compression Attached Memory Module 2 (CAMM2) connectors, and Argosy has LPDDR5 SOCAMM2 samples 3. CAMM2 parts can carry 644 pins with 1.00 x 1.38 mm pitch (the spacing between pins) 4 and use screw-lock attachment for vibration resistance 5.
Server integrators or OEMs not named in the article should sign connector partners now. Detachable modules need mechanical retainers and thermal frames (metal structures that spread heat and interface to heatsinks or cold plates), which lock into supply deals once a design is frozen.
Investors can track spending on LPDDR5X-compatible thermal solutions and high-density connector capacity, since even with the over-55% power reduction Samsung claims versus Registered Dual In-line Memory Modules (RDIMMs) compact modules still need advanced cooling.