As AI workloads continue to reshape data center design, memory architectures are evolving to meet rising demands for bandwidth, density, and power efficiency. One of the technologies gaining attention is SOCAMM2, a new standardized memory format aimed squarely at next-generation AI platforms.
Samsung introduces AI-focused SOCAMM2 memory
Samsung recently revealed a SOCAMM2 module built on LPDDR5, targeting high-performance AI data center systems. SOCAMM2 represents a shift away from traditional server DIMMs, offering a new physical layout designed to deliver more throughput in a smaller footprint.
The concept traces back to CAMM (Compression Attached Memory Module), which Dell originally created for thin laptops. That design was later handed over to JEDEC, allowing CAMM2 to evolve into an open industry standard. SOCAMM2 marks the first server-class implementation of that standardized approach.
LPDDR5 performance with lower energy use
At its core, SOCAMM2 relies on LPDDR5, a memory technology known for combining high bandwidth with low power draw. While LPDDR has long been associated with mobile devices, SOCAMM2 adapts it for data center environments, delivering performance comparable to DDR while consuming significantly less energy.
Samsung reports that SOCAMM2 can provide up to double the bandwidth of DDR5 RDIMMs commonly used in servers, with noticeably lower power requirements. Industry estimates place performance gains between 1.5x and 2x, while power usage may drop by more than half compared to standard DDR5.
Smaller modules, higher density
A key advantage of SOCAMM2 is its compact physical design. The modules use advanced stacking techniques that layer multiple memory dies within a single package, dramatically increasing density. This allows equivalent—or greater—amounts of memory to occupy far less board space than traditional DIMMs.
Because of this flexibility, SOCAMM2 can either complement conventional DDR memory or replace it entirely as system RAM, depending on platform design and workload needs.
Industry backing and standardization
SOCAMM2 is not tied to a single vendor’s ecosystem. After Dell collaborated with partners to define the original CAMM specification, the technology was formally adopted by JEDEC. The standards body expanded the design to include ECC and enterprise-grade reliability features, making it suitable for mission-critical environments.
According to Jim Handy, president of Objective Analysis, SOCAMM addresses real architectural constraints rather than introducing unnecessary complexity. He notes strong support from server CPU designers and Nvidia, driven by SOCAMM’s ability to deliver higher memory bandwidth, increased density, and reduced power consumption within a tight physical space.
Cost expectations and vendor support
Although stacked memory typically raises concerns about manufacturing costs, Handy argues that SOCAMM2 should not command a premium. Memory suppliers already produce stacked configurations at scale using packaging techniques similar to those employed in NAND flash, keeping pricing in line with traditional DRAM solutions.
In addition to Samsung, SK Hynix has confirmed plans to support SOCAMM2, though its timeline appears to trail competitors such as Samsung and Micron. Broader adoption is expected to accelerate alongside Nvidia’s Vera Rubin platform, currently anticipated to debut in Q2 2026