Fabless semiconductor company Xcena (엑시나), which develops next-generation memory infrastructure, took part in OCP Global Summit 2026 held at the San Jose Convention Center in the United States from Oct. 12 to 15 local time and showcased memory infrastructure technology for efficient AI inference, it said on Thursday.

Xcena presented two AI inference demos using its CXL (Compute Express Link)-based memory solution MX1 at the event.

First, it presented a demo in which multiple AI inference servers share a KV cache through a CXL memory pool. KV cache is data that AI stores from previously processed information and reuses to generate answers. Instead of storing KV cache separately on each server, saving it in a CXL memory pool and sharing it across multiple servers can reduce duplicate storage and the burden of data movement between servers, the company said.

Xcena also presented a demo that accelerates vector search, one of the key processes in retrieval-augmented generation (RAG), using the MX1 near-data processing (NDP) function.

Xcena is also working with SK hynix on technical cooperation to apply a CXL hybrid architecture to AI workloads.

Xcena CEO Jin-young Kim (김진영) said, "To increase AI infrastructure performance and efficiency, it is important not only to improve compute performance but also to use memory efficiently and reduce data movement." He added, "At this event, we showed how MX1 can be used in real AI workloads, and we will continue to work with partners across the ecosystem to advance next-generation AI memory infrastructure."