Breaking Through AI Data Bottlenecks... SK hynix Unveils Next-Generation Memory Solutions at 'OCP 2026'

Solutions for Server Architectures in the AI Era
Showcases Technology, Including a Memory-Centric AI Rack Demonstration
Exhibits HBM4, SOCAMM2, HBF and More

SK hynix has unveiled next-generation memory solutions designed to fundamentally address the “data bottleneck” caused by the explosive growth in artificial intelligence (AI) data.

A view of the SK hynix booth at the OCP Global Summit 2026, a data center hardware technology conference held in San Jose, California, from the 12th (local time) through the 15th. SK hynix

SK hynix said it showcased a range of next-generation AI memory solutions at OCP Global Summit 2026, a data center hardware technology conference held in San Jose, California, from October 12 (local time) through October 15. For the exhibition, SK hynix built a booth featuring a sturdy gray frame and neon-blue lighting to visualize “reliability” and “efficiency.” Divided into a “Memory Section” at the front and a “System Demo” section at the back, the booth demonstrated a new server system in operation and showcased the company’s technology leadership.

The System Demo section presented a series of key solutions to overcome power waste and data-processing limitations arising as AI services become more advanced. CMM (CXL Memory Module)-DDR5 is a memory solution based on the Compute Express Link (CXL) interface standard that allows multiple servers to share memory as needed. The second-generation 256GB product based on CXL 3.2, unveiled at the event, offers more than double the bandwidth of the previous generation and approximately 30% greater energy efficiency. SK hynix also presented results demonstrating, at the proof-of-concept (PoC) level, the effectiveness of KV (key-value) cache sharing in agentic AI environments and embedding sharing for the multimodal service OmniServe. The work was based on a memorandum of understanding (MOU) signed with NAVER Cloud in September last year.

The Memory-centric AI Rack System (MARS) integrates partner solutions into a single rack, with memory at its center, to maximize data-transfer efficiency. Using CXL pooled memory to directly connect graphics processing units (GPUs) and servers, it significantly improves KV cache reuse speed and system performance compared with conventional network communications. SK hynix validated the system’s ability to efficiently store and reuse continuously generated, large-scale KV caches using a hybrid architecture combining DRAM and solid-state drives (SSDs). Streaming based on pooled memory accelerates data transfer between AI models, dramatically reducing service response latency.

CMM-Ax is a next-generation memory that performs computations directly within memory, substantially reducing the power and cost of transferring data to GPUs. At the exhibition, SK hynix demonstrated that it can accelerate long-context computations that are difficult to handle with conventional GPU-based systems. The research was accepted for MICRO 2026, a leading conference in computer architecture, in recognition of its technological merit.

Visitors look around the "System Demo" section at SK hynix's booth at OCP Global Summit 2026, a data center hardware technology conference held in San Jose, California, from the 12th to the 15th (local time). SK hynix

The Memory Section at the front of the exhibition hall featured both SK hynix’s flagship lineup, which currently leads the memory market, and its next-generation products. In the enterprise SSD (eSSD) segment, the company showcased a customer-tailored lineup that included the high-performance PS1010 U.2; the PS1110 E3.S, which is under development; the high-capacity PS1101 E3.S; and the data center products PEB210 E1.S and PEB210 M.2 2280, which support direct liquid cooling (DLC). Its server DRAM lineup included the industry’s highest-capacity 256GB 3DS server DDR5 registered DIMM (RDIMM), 96GB and 64GB DDR5 RDIMMs, a 128GB DDR5 multiplexed rank DIMM (MRDIMM), and a 192GB SOCAMM2 module designed to improve power and space efficiency.

SK hynix also displayed a model of NVIDIA’s next-generation Vera Rubin Superchip and set up an interactive area where visitors could install high-bandwidth memory (HBM)4 and SOCAMM2 in a server model. Through an HBM model enlarged to one million times its actual size and accompanying monitors, the company highlighted the thermal-management and warpage-control performance of its key manufacturing technologies: through-silicon via (TSV) and Advanced Mass Reflow Molded Underfill (MR-MUF). It also displayed and compared actual HBM3, HBM4, and HBM4E products from different generations.

High-bandwidth flash (HBF) is a next-generation solution that applies TSV technology to NAND flash, adding high bandwidth to the large-capacity characteristics of conventional storage. SK hynix presented a tiered memory strategy: using HBM in environments that require fast response times, and HBF in AI inference environments that need large-capacity memory to serve many users.

SK hynix also shared its technology recommendations during presentation sessions held throughout the event. In a presentation on MARS, Kim Hosik, vice president in charge of memory system research, noted that the conventional CPU- and GPU-centric architecture faces the limits of the “memory wall” due to power consumption and bottlenecks. He emphasized that MARS, based on CXL technology, would fundamentally address KV cache reuse in agentic AI, accelerate responses from multimodal omni models, and resolve irregular memory-access issues in graph neural network (GNN) training.

Joo Youngpyo, vice president in charge of memory architecture, stressed the importance of software frameworks that connect customers with products across the ecosystem, as well as the need for standardization. Kim Jongryul, head of the AI System Infrastructure Team, introduced shared-memory-based token sharing and KV cache sharing technologies to improve agentic AI performance.

“This OCP Global Summit was significant because it gave us the opportunity to show directly how next-generation memory solutions operate in real systems,” SK hynix said. “We will continue to drive technological innovation so that AI technology can continue to advance with memory at its core.”

This content was produced with the assistance of AI translation services.