SK hynix presented next-generation memory solutions for AI infrastructure at OCP Global Summit 2026, including a 256GB CMM-DDR5 module based on CXL 3.2. The event took place October 12-15 in San Jose, California.
The company said the growing volume of data handled by AI has created data bottlenecks. At the summit, SK hynix also described conventional CPU- and GPU-centric architectures as facing a “memory wall” and sharply rising power consumption as generative AI and hyperscale data centers grow.
CMM-DDR5 lets multiple servers share a memory pool and use the capacity they need, based on the CXL interface standard. SK hynix said its second-generation product, optimized for AI workloads, more than doubles bandwidth to 80GB/s compared with the previous generation and improves energy efficiency by about 30%. The product also applies CXL Pooled functionality.
SK hynix demonstrated MARS (Memory Centric AI Rack System), which integrates partners’ solutions into a rack with memory at its core. In one demonstration with Arcee, Liqid and Marvell, multiple servers and GPUs used pooled memory to transfer or reuse KV cache; the company said this improved system performance compared with methods relying on network communication. A separate demonstration with AMD, Penguin, Xcena and Arcee used CXL memory and CXL hybrid memory to store and reuse KV caches, including loading required caches into DRAM in advance and storing the full cache on an SSD.
The company also demonstrated CMM-Ax, a memory solution that combines computation functions with memory. Its “Efficient Long Context LLM Serving” demonstration showed how CMM-Ax can accelerate long contexts that have been difficult to process on conventional GPU-based systems. SK hynix said the research was published to MICRO 2026.
In a presentation on MARS, Vice President Hosik Kim, head of the Memory Systems Research department, said: “This system implements rack-level memory pooling and dynamic zero-copy data sharing by connecting CPUs and GPUs with next-generation CXL technology, and it is structured to significantly increase the speed of data movement in high-load workloads.”
The exhibition also featured SK hynix’s eSSD and server DRAM products, HBM4E, HBM4 and HBM3, and a model of HBF (High Bandwidth Flash), a next-generation NAND solution that applies TSV technology to combine high capacity with high bandwidth. SK hynix said it will continue pursuing technological innovation so AI technology can advance with memory at its core.
