Samsung Announces SOCAMM2 Memory for Next-Gen AI Servers

Samsung has announced SOCAMM2, its LPDDR-based server memory module for AI data centers. The new memory module delivers high bandwidth and improved power efficiency to support the heavy computational workloads. The company says it has already supplied samples to its customer, likely referring to Nvidia.

SOCAMM2 offers high bandwidth and low power consumption for AI servers

Built on LPDDR5X DRAM, SOCAMM2 (Small Outline Compression Attached Memory Module) brings the strengths of LPDDR and modular architectures. Traditional DDR-based memory modules, such as RDIMM, remain common in standard servers, but SOCAMM2 is designed for the next-generation AI-accelerated servers. Samsung says the new module offers more than double the bandwidth of RDIMM while consuming over 55% less power. This makes it a solid option for energy-efficient, performance-driven AI servers.

SOCAMM2 features a detachable structure, making it easier to replace or upgrade compared to traditional soldered LPDDR memory. This design helps data center operators reduce maintenance time and system downtime. Thanks to the module’s improved power efficiency, it also helps reduce heat generation in AI servers. This allows data centers to improve cooling efficiency while better supporting high-density server deployments. In addition, the module uses a horizontal layout to improve system-level space utilization.

Samsung says it is working with Nvidia to optimize SOCAMM2 for the GPU maker’s AI systems. “Our ongoing technical cooperation with Samsung is focused on optimizing memory solutions like SOCAMM2 to deliver the high responsiveness and efficiency essential for AI infrastructure,” said Dion Harris, senior director, HPC and AI Infrastructure Solutions, Nvidia. The company is also contributing to industry-standardization efforts through JEDEC.

Industry insiders say Nvidia may need around 20 billion gigabytes of SOCAMM2 from global DRAM suppliers next year. Samsung will likely supply around half of that volume (about 10 billion gigabytes). To meet this demand, the Korean firm will need to make between 30,000 and 40,000 wafers per month.