Meta and Panmnesia propose CXL-based AI data center architecture

Researchers from KAIST, the startup Panmnesia, and Meta have proposed a next-generation AI data center architecture designed to function as a single, tightly-coupled computing system rather than a collection of independent servers.

Published in Nature Reviews Electrical Engineering, the proposed model utilizes Compute Express Link (CXL) technology to bridge CPUs, AI accelerators, and memory. This approach aims to create a single coherence domain, effectively allowing a data center to operate with the coordination of a single large chip.

Key technical advantages include the ability to connect up to 960 AI accelerators within a single connection domain, which is approximately 13 times the capacity of current NVLink-based racks. The architecture also seeks to significantly reduce data transfer latency, moving from microsecond scales to nanosecond scales, by reducing the traditional boundaries between individual server resources.