Chuangxin Cloud

All Categories

Language

Home > The Blogs

SK Hynix has officially begun mass production of 192GB SOCAMM2

4/22/2026 1:36:09 AM

The company has officially begun mass production of 192GB high-capacity products, designed for the NVIDIA Vera Rubin platform.
Featuring highly integrated DRAM based on the 1c-nanometer process, it achieves maximum energy efficiency.
"Through close collaboration with NVIDIA, we aim to alleviate bottlenecks in AI infrastructure and deliver optimal performance."

On April 20, 2026, SK Hynix announced the official mass production of its 192GB (Gigabyte) capacity SOCAMM2 product, based on sixth-generation 10-nanometer-class (1c) LPDDR5X low-power DRAM.

SOCAMM2 is a memory module optimized for server environments, primarily derived from low-power memory designed for mobile devices like smartphones. Its main application targets next-generation AI servers and similar use cases.

SK Hynix stated, "Our SOCAMM2, manufactured using the 1c process, is a solution specifically optimized for high-performance AI computing. Compared to traditional RDIMM, it offers over twice the bandwidth and reduces power consumption by more than 75%."

The company specifically emphasized that this SOCAMM2 product is designed for NVIDIA's Vera Rubin platform.

Furthermore, the product can fundamentally alleviate the storage bottleneck faced by large AI models with hundreds of billions of parameters during training and inference. SK Hynix expects it to significantly enhance overall system processing speeds.

The company also stated: "As AI development shifts from training to the inference stage, SOCAMM2-which supports the low-power operation of large language models-is gaining significant attention as a next-generation memory solution. To meet the demands of global Cloud Service Provider (CSP) customers, the company has proactively established a stable mass production system."

SK Hynix's CMO, Jin Ju-seon, stated: "With the supply of the 192GB SOCAMM2, we have established a new AI memory performance benchmark. We will collaborate closely with global AI clients to become 'the most trusted AI memory solutions provider.'"

*** SOCAMM2 (Small Outline Compression Attached Memory Module 2): A low-power DRAM module featuring a slim profile and scalability, optimized for AI servers. It uses a compression connector to enhance signal integrity and allow for easy replacement.

*** RDIMM (Registered Dual In-line Memory Module): A server/workstation DRAM module that inserts a register or clock buffer between the memory controller and DRAM chips to relay address/command signals.

PREV：AI Era High Pressure Closed Loop How Chip Giants Face the Dual Narrative of Costs and Expansion

NEXT：Computing-Power Synergy Drives AIDC Energy Storage Boom Component Value Restructuring is Underway