SK hynix has started mass production of 192 GB SOCAMM2 modules for NVIDIA Vera Rubin

After Micron, the South Korean company SK hynix has announced the start of mass production of 192 GB SOCAMM2 modules, based on LPDDR5X memory manufactured using a 1cnm process (sixth generation of the 10 nm class). This is a solution that marks a significant evolution in the adoption of low-power memory in server environments.

The SOCAMM2 modules (Small Outline Compression Attached Memory Module 2) arise from the adaptation of LPDDR memory—typically used in mobile devices—for data center environments, aiming to offer a more efficient alternative compared to traditional server modules.

The new modules are specifically developed for the NVIDIA Vera Rubin platform, one of the next-generation architectures intended for large-scale AI workloads.

From a technical standpoint, SOCAMM2 introduces a compact form factor with greater scalability, a compression connector that enhances signal integrity, and ensures simplified module replacement.

According to the company, the new solution is capable of providing more than double the bandwidth compared to traditional RDIMM modules, accompanied by an improvement in energy efficiency exceeding 75%.

One of the most relevant aspects concerns the impact on the performance of large models. SK hynix emphasizes that SOCAMM2 is designed to reduce memory bottlenecks during training and inference of LLMs with hundreds of billions of parameters.

"With the launch of the 192 GB SOCAMM2, SK hynix has set a new standard in memory performance for artificial intelligence," said Justin Kim, President and Head of the AI Infra division (CMO, Chief Marketing Officer) at SK hynix. "We will consolidate our position as the most reliable memory solutions provider for artificial intelligence, thanks to close collaboration with our global customers in the AI sector."