XCENA Inc., a startup developing a memory device to accelerate artificial intelligence clusters, has raised $135 million in Series B funding. The round was led by South Korean funds Atinum Investment and IMM Investment, with contributions from more than half a dozen other institutional backers. The company is now valued at $570 million.
Founded in 2022 by former employees of Samsung Electronics and SK hynix, XCENA’s flagship product is the MX1, a computational memory controller designed to speed up data management for AI inference workloads.
Large language models (LLMs) rely on a data structure called a KV cache to process user prompts. When that cache cannot fit into a graphics card’s onboard memory, it must be offloaded to slower external DRAM, creating delays. The MX1 addresses this by combining up to two terabytes of DRAM with thousands of CPU cores, eliminating those bottlenecks and boosting inference performance.
The chip also reduces the need for duplicate calculations by allowing the KV cache to be reused across user requests, further reducing processing overhead. Additionally, the MX1 can accelerate analytics workloads like Apache Spark by minimizing data travel times between memory and processing cores.
The MX1’s CPU cores are based on the open-source RISC-V architecture and are organized into four-core clusters with dedicated L1 cache. XCENA provides APIs that allow developers to port their workloads without major code changes, alongside a simulation tool for reliability testing.
The company plans to manufacture the MX1 using Samsung’s four-nanometer process, with mass production slated to begin by the end of the year and revenue expected in 2027. XCENA will use the new capital to develop additional computational memory products and accelerate partnerships with hyperscalers.