SK hynix Begins Mass Production of 192GB SOCAMM2 'Setting a New Standard for AI Server Memory Performance'

SK hynix Inc. began mass production of the 192GB SOCAMM2 memory module standard, using a 1cnm process. The company said the move targets next-generation AI server memory with a low-power design.

SK hynix described SOCAMM2 as a way to address memory bottlenecks during training and inference of large language models with hundreds of billions of parameters. It said the product is intended as a primary memory solution for next-generation AI servers and that supply and mass production capacity were stabilized to meet demand from global Cloud Service Provider customers.

SOCAMM2 uses low-power DRAM based on LPDDR5X and repurposes low-power memory previously used mainly in mobile products for server environments. SK hynix said its 1cnm-based SOCAMM2 provides more than double bandwidth and over 75% improved power efficiency compared to conventional RDIMM, which includes a register or buffer chip to relay address and command signals.

The company said it designed SOCAMM2 to support NVIDIA Vera Rubin and that it would collaborate closely with NVIDIA to solve bottlenecks in AI infrastructure and provide optimal performance. “By supplying the 192GB SOCAMM2, SK hynix has established a new standard for AI memory performan3e,” Just3ein Kim, President & Head of AI Infra (CMO, Chief Marketing Officer) at SK hynix said. “We will solidify our position as the most trusted AI memory solution provider, through close collaboration with our global AI customers.”

SK hynix expects SOCAMM2 to resolve memory bottlenecks in large language model training and inference.