[Edaily Reporter Choi Oh-hyun ] SK hynix has unveiled next-generation memory technology designed to resolve data processing bottlenecks in artificial intelligence (AI) data centers. The company is accelerating the development of technologies—beyond individual semiconductor products such as High Bandwidth Memory (HBM)—that enhance the overall performance and power efficiency of servers by focusing on memory.

SK hynix announced on the 16th that it participated in the “OCP Global Summit 2026,” held in San Jose, California, from the 12th to the 15th, where it showcased next-generation memory solutions and server systems.

The OCP Global Summit is an event for sharing data center hardware designs and technical standards. At this event, SK hynix focused on introducing memory sharing technology utilizing Compute Express Link (CXL), a next-generation interface.

In the “System Demo” section, three solutions—△CMM-DDR5, △MARS, and △CMM-Ax—were introduced in succession. “CMM (CXL Memory Module)-DDR5” is a memory solution that improves resource utilization efficiency by sharing a memory pool that multiple servers can divide and use as needed, based on the CXL interface standard.

At this exhibition, a second-generation product (256GB) optimized for AI workloads based on CXL 3.2 was also unveiled. Compared to the previous generation, its bandwidth has more than doubled (80GB/s), and energy efficiency has improved by approximately 30%. CXL is a next-generation solution that efficiently connects CPUs, GPUs, and memory within computing systems to support large-scale, ultra-high-speed computations.

SK hynix also unveiled CXL-based AI service technology currently under joint development with Naver Cloud. This approach accelerates data processing by enabling multiple servers to share a “KV cache,” where AI stores previous computation results. The two companies confirmed the technology’s feasibility through a proof of concept (PoC) utilizing an actual AI service environment.

The company also introduced “MARS,” a memory-centric AI rack system. This system is designed so that servers and graphics processing units (GPUs) share memory resources. At the event, SK Hynix, together with partners such as AMD and Marvell, demonstrated technologies that enhance the data processing speed and resource utilization efficiency of AI agents.

The company also unveiled “CMM-Ax,” which performs computations directly within memory. This technology reduces power consumption and processing latency by minimizing the process of transferring data to the GPU.

The company also introduced its next-generation product lineup. SK hynix exhibited HBM4 and HBM4E alongside a model of NVIDIA’s “Vera Rubin” superchip. It also showcased “HBF,” a next-generation NAND technology that, like HBM, stacks semiconductors vertically to achieve high capacity and fast data transfer speeds.

Kim Ho-sik, Executive Vice President of Memory System Research at SK hynix, pointed out in his presentation that existing CPU- and GPU-centric architectures are facing bottlenecks—known as the “Memory Wall”—and surging power consumption. He then proposed a “memory-centric AI rack system” as the solution.

“This system utilizes next-generation CXL technology to implement rack-level memory pooling and dynamic zero-copy data sharing between CPUs and GPUs,” explained Vice President Kim. “It features an architecture capable of significantly boosting data transfer speeds for high-load workloads.” Memory pooling is a technology that consolidates memory resources into a single pool and allocates them to individual servers as needed, while zero-copy is a technology that shares data directly without separate copying.

SK hynix stated, “At this OCP Global Summit, we were able to demonstrate firsthand how our next-generation memory solutions operate in actual systems,” adding, “We will continue to drive technological innovation to ensure that AI technology continues to advance, centered on memory.”