XCENA MX1 production lineup launched for hyperscale AI infrastructure

XCENA, a company focused on memory-centric computing for AI infrastructure, has introduced MX1, its new production memory lineup designed to address the capacity, utilization, and data-movement bottlenecks increasingly limiting AI inference at scale. The company is showcasing MX1 at the Future of Memory and Storage conference, running August 4-6 in Santa Clara.

 

The product responds to a growing challenge in AI infrastructure: as generative AI models get larger, work with longer context windows, and generate increasingly large KV cache footprints, meaning the temporary memory used to store intermediate computation results during inference, memory itself is becoming the main constraint on system performance. High-bandwidth memory remains expensive and limited in capacity, while simply adding more servers to scale out often leaves DRAM underused on some machines while driving up overall costs. XCENA built its CXL-based MX1 lineup, using the Compute Express Link standard that lets memory be shared flexibly across processors, to give operators an alternative: scaling memory as efficiently as they already scale computing power. Paired with Intel's Xeon 6 platforms, MX1 is meant to demonstrate how CXL-based memory can help address AI workloads that demand large amounts of memory.

 

Jin Kim, CEO of XCENA, said AI performance ...