As artificial intelligence workloads continue to reshape data center design, a new memory format is gaining attention for its ability to deliver higher bandwidth in a smaller, more energy-efficient package. Known as SOCAMM, the technology is increasingly viewed as a strong fit for next-generation AI infrastructure.

Momentum around SOCAMM accelerated last month when Samsung introduced SOCAMM2, a memory module built on LPDDR5 and optimized specifically for AI data center platforms. SOCAMM2 represents the first standardized generation of CAMM-based memory developed through industry collaboration rather than a single vendor.

The underlying concept originates from CAMM (Compression Attached Memory Module), a form factor initially created by Dell for use in laptops. Dell later handed the specification to the JEDEC standards body, paving the way for broader adoption. SOCAMM2 is the enterprise- and server-oriented evolution of that idea.

Unlike traditional server memory, SOCAMM2 relies on LPDDR5, a memory technology commonly associated with mobile devices due to its high bandwidth and low power draw. By adapting LPDDR5 for servers, SOCAMM2 delivers performance comparable to DDR-class memory while significantly reducing energy consumption.

According to Samsung, SOCAMM2 can deliver up to double the bandwidth of standard DDR5 RDIMM modules used in servers, while consuming substantially less power. Independent estimates place the performance gain between 1.5× and 2×, with power usage reduced by as much as 55% compared to conventional DDR5.

Physical design is another key advantage. SOCAMM2 modules are notably smaller than standard DIMMs because they use stacked memory chips, allowing multiple layers of DRAM to be combined into a single package. This density enables higher memory capacity in a much smaller footprint, freeing up valuable motherboard space. SOCAMM2 can function either alongside traditional DDR memory or as the primary system memory.

Because the CAMM specification was transferred to JEDEC, SOCAMM2 is widely regarded as an open industry standard rather than a proprietary solution. JEDEC has expanded the original design to include enterprise-grade features such as ECC and advanced error correction, making it suitable for data center and mission-critical workloads.

Industry analysts emphasize that SOCAMM is solving real-world challenges rather than chasing novelty. Jim Handy, president of Objective Analysis, notes that strong support from CPU vendors and Nvidia stems from SOCAMM’s ability to provide faster memory interfaces, higher density, and lower power consumption—critical factors for AI systems.

While stacked memory designs are often assumed to be more expensive, Handy argues that cost should not be a barrier. Memory manufacturers already sell stacked configurations at prices comparable to conventional DRAM, using mature packaging techniques similar to those employed in NAND flash production.

Other major players are also moving toward adoption. SK Hynix has confirmed plans to support SOCAMM2, though it has yet to announce a timeline and is believed to be trailing Samsung and Micron. Industry expectations point to a broader SOCAMM2 rollout around the second quarter of 2026, coinciding with Nvidia’s upcoming Vera Rubin AI platform launch.

With rising AI memory demands and increasing pressure to reduce power consumption, SOCAMM2 is positioning itself as a compelling alternative to traditional server memory—and a likely fixture in future AI data center architectures.