SK hynix said Monday it has begun mass production of SOCAMM2, a next-generation memory module designed to significantly boost performance and power efficiency in artificial intelligence servers.
The new Small Outline Compression Attached Memory Module 2 (SOCAMM2) has been optimized for Nvidia’s upcoming Vera Rubin platform, highlighting deepening collaboration between the two companies.
SOCAMM2 integrates 192 gigabytes of memory using sixth-generation 10-nanometer (1c) LPDDR5X DRAM, a low-power memory typically used in mobile devices. Unlike conventional server modules based on ordinary DDR5, the design vertically stacks LPDDR chips to improve energy efficiency while maintaining high performance.
“We expect SOCAMM2 to fundamentally address memory bottlenecks in training and inference for large language models with hundreds of billions of parameters, significantly accelerating overall system performance,” SK hynix said.
According to the company, SOCAMM2 delivers more than twice the bandwidth and over 75 percent greater power efficiency compared with conventional DDR5 RDIMM modules, making it well suited for high-performance AI workloads.
Performance has also improved from the previous generation, with data transfer speeds increasing to 9.6 gigabits per second from 8.5 Gbps in SOCAMM1. The module is designed with a higher number of input/output pins — contact points that enable data transfer — compared to traditional DRAM, further enhancing throughput.
These gains in speed and efficiency are expected to lower total cost of ownership for hyperscale data center operators, where investment decisions are driven not only by GPU or HBM costs but also by rack-level performance, power consumption, cooling requirements and operational efficiency.
While SOCAMM does not match the ultra-high bandwidth of high bandwidth memory, its LPDDR-based architecture allows for a simpler manufacturing process and higher yields, offering a cost advantage on a per-capacity basis, an industry official said.
The modular form factor also marks a departure from conventional LPDDR memory, which is typically soldered onto boards and cannot be replaced, enabling greater flexibility in system design and maintenance, the company added.
SK hynix said it worked closely with Nvidia to tailor SOCAMM2 for the Vera Rubin platform, slated for launch in the second half of the year. The chipmaker is also expected to supply its next-generation HBM4 memory for the platform.
It said it expected the use of SOCAMM to allow for multitier structures comprising high bandwidth memory, SOCAMM, DDR5 RDIMM and CXL-based expansion memory, according to industry officials.
“In this hierarchy, SOCAMM serves as an intermediate layer, handling frequently accessed ‘hot’ data and buffering workloads between HBM and system memory to reduce bottlenecks,” an official said.
SK hynix President and head of AI Infra Kim Ju-seon said the 192GB SOCAMM2 sets a new benchmark for AI memory performance.
“We will strengthen our position as a trusted AI memory solutions provider through close collaboration with global AI customers,” Kim said.
herim@heraldcorp.com
