Panmnesia, a South Korea-based fabless semiconductor startup, said it collaborated with Meta on a Compute Express Link (CXL)-based AI datacenter architecture described in an invited Review in Nature Reviews Electrical Engineering.

The companies said the proposed design would let a datacenter operate like a single chip.

Compared with Nvidia’s rack-scale GB200 NVL72 reference platform, the architecture would expand the coherence domain to up to about 13x as many accelerators and cut cross-rack access latency from the microsecond range to several hundred nanoseconds.

Panmnesia said it has implemented and validated core parts of the architecture in silicon and is preparing them for commercial supply.

The Review also discusses hardware-based security controls and CXL-over-optics to extend system reach beyond the limits of electrical signaling.

CXL enables pooling and memory expansion in servers, while Nvidia’s scale-up interconnect approach is optimized for tightly coupled graphics processing unit systems.

The collaboration fits Meta’s broader rack-scale infrastructure work through the Open Compute Project.

AMD has said its Helios AI rack is built on Meta’s 2025 Open Compute Project design.

🔗 Source: Panmnesia Inc.

Recent Panmnesia developments