SK hynix begins mass production of 1c SOCAMM2 AI server chips
SK hynix's 192GB small outline compression attached memory module 2 (SOCAMM2) / Courtesy of SK hynix
SK hynix said Monday it has begun mass production of its 192GB small outline compression attached memory module 2 (SOCAMM2), a next‑generation chip for artificial intelligence (AI) servers produced with its sixth-generation 10‑nanometer‑class (1c) process.
SOCAMM2 is a server memory module that leverages low-power double data rate (LPDDR) memory chips commonly used for smartphones, aimed at cutting power consumption to roughly one-third of conventional server modules.
Designed specifically for AI servers, it uses a thin, high-density form factor, improving signal integrity and making it easier to swap or upgrade.
SOCAMM2 is gaining attention among AI data center operators as power efficiency becomes increasingly important for managing total cost of ownership.
While commonly used AI memory such as high-bandwidth memory (HBM) is mounted within the package of logic chips such as graphics processing units (GPUs) or central processing units (CPUs), SOCAMM2 is typically placed next to the logic chips on the system board. In this setup, HBM supports computing acceleration, while SOCAMM2 improves overall system-level power efficiency by complementing conventional DDR-based memory modules.
Notably, SK hynix uses the 1c process for manufacturing LPDDR5X memory for the SOCAMM2.
Of the 1a, 1b and 1c process generations, 1c is considered one of the most advanced nodes currently available, delivering both performance gains and improved power efficiency. Industry officials said DDR5 built on the 1c process is known to offer about 11 percent faster speeds and more than 9 percent better power efficiency compared with 1b-based DDR5.
“With its 1c process, SOCAMM2 delivers more than twice the bandwidth of conventional RDIMMs (registered dual in-line memory modules) while improving energy efficiency by over 75 percent, making it a solution tuned for high-performance AI workloads,” the company said.
The company noted that the product has been optimized for Nvidia’s Vera Rubin, a next-generation AI computing platform.
SK hynix expects the new module to significantly ease memory bottlenecks, wherein data delivery falls behind GPU processing speeds, during the training and inference of large-scale AI models with hundreds of billions of parameters.
Parameters are learned values that a model uses to analyze data and are widely regarded as an indicator of the model’s scale or size.
When SOCAMM2 is introduced, the memory hierarchy in AI servers will become a multitier structure consisting of HBM, SOCAMM, DDR5 memory module and compute express link memory serving as expanded memory.
Nvidia's Vera Rubin platform is displayed at SK hynix's booth at GTC 2026 in San Jose, March 16. Nvidia CEO Jensen Huang signed his name on a processor package in the upper right. Yonhap
“As the AI market shifts from training to inference, SOCAMM2 is gaining traction as a next-generation memory solution capable of running large language models with high energy efficiency,” the company said, adding that it has stabilized mass production quickly to meet demand from global cloud service providers.
“The company has set a new standard for AI memory performance with the launch of the 192GB SOCAMM2," said Kim Ju-seon, chief marketing officer of SK hynix. "Through close collaboration with global AI customers, we will strengthen our position as a trusted AI memory solutions provider.”