SK hynix announced Monday the commencement of mass production for SOCAMM2, its groundbreaking next-generation memory module. This innovative module is engineered to dramatically enhance both performance and power efficiency, specifically targeting advanced artificial intelligence (AI) servers.
This advanced Small Outline Compression Attached Memory Module 2 (SOCAMM2) has been meticulously optimized for integration with Nvidia’s highly anticipated Vera Rubin platform. This optimization underscores the deepening strategic collaboration between SK hynix and Nvidia, focusing on future AI infrastructure.
The SOCAMM2 module boasts an impressive 192 gigabytes of memory capacity, leveraging SK hynix’s sixth-generation 10-nanometer (1c) LPDDR5X DRAM. While LPDDR5X is often associated with low-power mobile devices, SOCAMM2’s unique design vertically stacks these LPDDR chips. This innovative architecture significantly improves energy efficiency while consistently delivering high performance for demanding AI applications, setting it apart from conventional DDR5 server modules.
SK hynix stated, “We anticipate SOCAMM2 will fundamentally resolve critical memory bottlenecks frequently encountered in the training and inference processes of large language models (LLMs) involving hundreds of billions of parameters. This will lead to a significant acceleration of overall system performance in AI deployments.”
The company further highlights that SOCAMM2 delivers over double the bandwidth and more than 75 percent greater power efficiency when compared to conventional DDR5 RDIMM modules. These superior characteristics make SOCAMM2 exceptionally well-suited for the most intensive high-performance AI workloads and data center environments.
Significant performance enhancements have also been achieved compared to its predecessor, SOCAMM1. Data transfer speeds have increased from 8.5 Gbps to an impressive 9.6 gigabits per second. Furthermore, the SOCAMM2 module incorporates a higher density of input/output (I/O) pins—critical contact points facilitating data transfer—beyond what traditional DRAM offers, thereby substantially boosting overall data throughput and efficiency.
These remarkable gains in speed, bandwidth, and power efficiency are poised to significantly lower the total cost of ownership (TCO) for hyperscale data center operators. For these critical AI infrastructures, investment decisions extend beyond just GPU or HBM costs, encompassing vital factors like rack-level performance, optimized power consumption, reduced cooling requirements, and enhanced overall operational efficiency.
While SOCAMM modules may not offer the ultra-high bandwidth characteristic of High Bandwidth Memory (HBM), an industry official noted that its LPDDR-based architecture provides distinct advantages. This includes a simpler manufacturing process and higher production yields, translating into a compelling cost advantage on a per-capacity basis for AI memory solutions.
The modular form factor of SOCAMM2 also represents a significant innovation, departing from conventional LPDDR memory which is typically soldered directly onto system boards, making replacement impossible. This new approach enables unparalleled flexibility in system design, upgrades, and maintenance, according to SK hynix.
SK hynix confirmed its close collaboration with Nvidia to precisely tailor SOCAMM2 for optimal performance within the Vera Rubin platform, which is slated for its highly anticipated launch in the latter half of the year. Furthermore, SK hynix is positioned to supply its cutting-edge next-generation HBM4 memory solutions for the same advanced platform.
Industry officials anticipate that the integration of SOCAMM will facilitate sophisticated multitier memory architectures. These advanced structures will dynamically comprise High Bandwidth Memory (HBM), SOCAMM, DDR5 RDIMM, and CXL-based expansion memory, creating a highly optimized memory hierarchy for AI.
An official elaborated, “Within this sophisticated memory hierarchy, SOCAMM uniquely functions as an intelligent intermediate layer. It adeptly handles frequently accessed ‘hot’ data and efficiently buffers workloads, strategically bridging the gap between ultra-fast HBM and broader system memory to significantly reduce critical bottlenecks.”
Kim Ju-seon, President and Head of AI Infra at SK hynix, affirmed that the 192GB SOCAMM2 module establishes a groundbreaking new benchmark for AI memory performance.
Kim concluded, “We are committed to solidifying our leadership as a trusted provider of cutting-edge AI memory solutions through sustained, close collaboration with our global AI customers, driving the future of artificial intelligence.”
