SK hynix (000660.KS) said Thursday it has begun mass production of its 192-gigabyte (GB) SOCAMM2 AI memory module, which will be mounted on Nvidia's latest artificial intelligence (AI) accelerator, Vera Rubin.
The product is the industry's first to be optimized for Nvidia's next-generation Vera Rubin accelerator. SK hynix expects SOCAMM2 to reduce memory bottlenecks that occur during the training and inference of hyperscale AI models and to improve system processing speeds.
SOCAMM2 is an AI-specialized memory module that adapts LPDDR, which has been primarily used in smartphones, for server environments. Unlike conventional LPDDR, which is soldered directly onto motherboards, SOCAMM2 uses a compression-type connector in a modular form that allows for easy attachment, detachment and replacement. When applied to servers, it maintains high performance while reducing power consumption.
The 192GB product is built on SK hynix's 10-nanometer-class sixth-generation DRAM (1c) process. According to the company, SOCAMM2 offers more than twice the bandwidth of RDIMM, the conventional server memory module, and improves energy efficiency by more than 75%. This helps ease memory bottlenecks in running hyperscale AI models and enhances overall system processing performance, the company said.
Demand has been growing in AI data centers for a new memory tier that can bridge the gap in performance and power efficiency between high-bandwidth memory (HBM), which provides ultra-high bandwidth, and capacity-focused DDR5 RDIMM. By combining the low-power characteristics of LPDDR with the maintenance convenience of a modular design, SOCAMM2 is emerging as a key pillar of AI infrastructure alongside HBM, DDR5 and Compute Express Link (CXL).
As the AI market shifts from training to inference, the importance of memory solutions capable of running large language models (LLMs) with low power consumption is growing further. SOCAMM2 is expected to contribute to reducing the total cost of ownership (TCO) of data centers by increasing the volume of AI computations that can be processed under the same power conditions.
SK hynix said it stabilized its mass production system early to meet demand from global cloud service providers (CSPs). The company's strategy is to extend the technological leadership it has secured in HBM to SOCAMM2, building a "full-stack" memory solution framework spanning the entire AI process from training to inference.
"With the supply of the 192GB SOCAMM2 product, we have set a new standard for AI memory performance," said Kim Ju-seon, president of AI Infra at SK hynix. "Based on close cooperation with global AI customers, we will establish ourselves as the most trusted AI memory solutions company."
