SK hynix Pitches Memory-Centric Servers With CXL Pooled Memory at OCP 2026

CMM-DDR5: a shared memory pool for multiple servers

The summit ran from the 12th to the 15th in San Jose, California. OCP Global Summit is a conference for sharing and standardizing the design and specifications of data center hardware. According to SK hynix, its server proposal drew considerable attention as AI infrastructure was a central theme this year.

The first product in the System Demo section was CMM-DDR5, a CXL Memory Module. It lets multiple servers share a memory pool and use as much of it as they need, based on the CXL interface standard. SK hynix unveiled a second-generation 256GB product optimized for AI workloads and based on CXL 3.2.

According to the company, the product more than doubles bandwidth (80GB/s) compared with the previous generation and improves energy efficiency by about 30%. CXL Pooled functionality is also applied.

SK hynix said it signed a business agreement with NAVER Cloud in September last year and has worked with it since on bottlenecks in cloud AI services. At the demo, it presented proof-of-concept results for a KV cache-sharing solution in an agentic AI environment and an embedding-sharing solution in OmniServe, an omni-modal service. The company said it analyzed real customer systems and workloads to confirm the applicability of its CXL-based solutions.

MARS: a memory-centric rack built with partners

MARS, short for Memory Centric AI Rack System, integrates partners' solutions into a single rack with memory at its core. SK hynix presented three demos with several partners.

• AI agent serving with CXL pooled memory: with Arcee, Liqid and Marvell, SK hynix showed a system connecting multiple servers and GPUs through pooled memory. The company said transferring or reusing a KV cache over CXL pooled memory significantly improves performance compared with conventional methods that rely on network communication.
• AI agent serving with CXL hybrid memory: with AMD, Penguin, Xcena and Arcee, it showed a memory rack using CXL memory and CXL hybrid memory, which combines DRAM and SSD. The rack can predict which KV caches will be needed from prior conversation and task history and load them into DRAM in advance. Storing the full KV cache on an SSD delivers fast reads and high capacity at once, the company said.
• Lower response latency: with Pickle, a personal AI agent service provider, it showed that streaming based on pooled memory can speed up data transfers between the multiple AI models a personal AI agent needs, so results reach users sooner.

CMM-Ax computes inside the memory

CMM-Ax combines computation functions with memory. By performing computations directly inside the memory, without moving data, it can reduce the power and cost burden on data centers, according to SK hynix.

In a demo of its Efficient Long Context LLM Serving inference system, the company said CMM-Ax can efficiently accelerate long contexts that have been difficult to process on conventional GPU-based systems. The underlying research was published at MICRO 2026, which SK hynix describes as the most prestigious academic conference in computer architecture.

Current flagship products on display

A separate Memory Section showed the products SK hynix says it currently leads the market with, including eSSDs and server DRAM for AI infrastructure. The eSSD lineup included the PS1010 U.2, the PS1110 E3.S and PS1101 E3.S, which are under sequential development, and the PEB210 E1.S and PEB210 M.2 2280, which support Direct Liquid Cooling.

For server DRAM, the company introduced a DDR5 RDIMM lineup led by the 256GB 3DS RDIMM, which it says delivers the industry's largest capacity for AI workloads, with 96GB and 64GB options also shown.

What it means

As agentic AI produces ever larger KV caches, SK hynix is positioning pooled and compute-capable memory as a way to avoid network and GPU limits. The results shown are demos and proof-of-concept validations, and the source gives no pricing or availability, so commercial rollout timing remains unclear.