According to industry sources on the 16th, XCENA demonstrated AI inference technologies using its Compute Express Link (CXL)-based memory solution, MX1, at the OCP Global Summit 2026, held in San Jose, California, from the 12th to the 15th. The demonstrations focused on technology for sharing key-value (KV) caches across multiple servers and accelerating vector searches for retrieval-augmented generation (RAG). A KV cache stores contextual information that an AI system has already processed so it can reuse that information when generating responses. When multiple inference servers each store the same information independently, memory consumption and the burden of transferring data between servers increase.
XCENA demonstrated an architecture in which KV caches are stored in a CXL memory pool and shared across multiple servers. CXL is an interface technology that connects processors, memory and other components. It enables memory to be expanded for use across multiple systems rather than being confined to individual servers. MX1 combines a CXL device that expands server memory capacity with near-memory processing capabilities. A single CXL Type 3 device supports up to 2 terabytes of DDR5 memory expansion, SSD-based capacity tiers, multiple RISC-V-based processing cores and vector engines. Rather than replacing CPUs or GPUs, the device is designed to handle workloads that involve frequent memory access.
XCENA is also validating CXL hybrid memory with SK hynix using its SSD-based expansion technology, InfiniteMemory. At the event, SK hynix showcased hybrid memory using the MX1 chip and an architecture based on remote direct memory access (RDMA). XCENA said the technology can provide greater capacity scalability than conventional DRAM-based CXL memory modules while delivering better performance than SSDs. XCENA CEO Jinyoung Kim said, “To improve the performance and efficiency of AI infrastructure, it is important not only to increase computing performance but also to use memory efficiently and reduce data movement.” He added, “We will continue working with partners across the ecosystem to advance next-generation AI memory infrastructure.”
