Panmnesia and Meta Propose CXL-Based AI Datacenter Architecture in Nature Journal

AI infrastructure startup Panmnesia and global hyperscaler Meta have jointly proposed a next-generation AI datacenter architecture, published by invitation in Nature Reviews Electrical Engineering. This architecture integrates CPUs, accelerators, and memory into a single datacenter-scale domain based on the Compute Express Link (CXL) standard. The proposed design aims to significantly enhance unified accelerator connection scale and reduce cross-rack access latency, addressing challenges in scaling AI models to trillion-parameter scales. This development is seen as a crucial advancement in the foundational hardware and interconnects for future AI.