Meta and Panmnesia propose CXL data center for 960 AI accelerators

Meta is working with Korean startup Panmnesia on a data center design intended to let thousands of processors operate as one coherent system. The proposal uses Compute Express Link (CXL) to connect CPUs, accelerators, and memory across multiple racks without conventional network links.

The architecture could place up to 960 AI accelerators in one coherence domain. It is designed to reduce variable communication delays that can make other accelerators wait during repeated AI-training stages. Ethernet and InfiniBand normally connect systems beyond individual servers, but packet processing and software coordination can increase latency variation as workloads expand.

Panmnesia's design uses a high-fan-out switch, a link acceleration unit, and a fabric controller arranged in trays, pods, and a fabric. The company says its fabric controller and link acceleration unit have completed silicon validation, while its switch has been fabricated.

The design compares with NVIDIA's GB200 NVL72, where one CPU coordinates two accelerators through NVLink-C2C. Panmnesia says one CPU could coordinate 16 accelerators, an eightfold increase. Around 60 such groups could form a domain with approximately 960 accelerators.

Cross-rack access could fall from microsecond-level timing to several hundred nanoseconds. Electrical CXL signalling reaches about seven meters at 128 GT/s with two retimers, so Panmnesia proposes optical CXL links for longer distances and says it has completed hardware proof-of-concept validation.

Panmnesia CEO Myoungsoo Jung said, “CXL enables the entire datacenter to operate as a single computing system.”