Meta is pushing the boundaries of artificial intelligence infrastructure by developing innovative networking technologies designed to support the growing demands of AI workloads. At the heart of this effort is the company's Disaggregated Scheduled Fabric (DSF) architecture, which has evolved into a 2-stage design capable of supporting up to 18,432 XPUs. This architecture enables non-blocking connectivity across entire data center buildings, making it a fundamental building block for constructing AI clusters that span regions.
Meta's AI infrastructure is also benefiting from the development of Non-Scheduled Fabric (NSF), a new architecture designed specifically for ultra-large AI clusters. NSF utilizes shallow-buffer Ethernet switches to achieve low round-trip latency and implements adaptive routing for dynamic load balancing and congestion mitigation. This approach allows Meta to build massive AI clusters that can handle the most demanding AI workloads.
In addition to its in-house developments, Meta is collaborating with industry leaders to advance AI networking standards. The company is a founding member of the Ethernet for Scale-Up Networking (ESUN) initiative, which brings together companies like AMD, Arista, and NVIDIA to develop Ethernet standards for scale-up AI networking. ESUN aims to adapt Ethernet's proven ecosystem to meet the stringent requirements of accelerator interconnects, enabling more efficient and scalable AI infrastructure.
Meta's commitment to innovation in AI infrastructure is also evident in its work on optical connectivity. The company has introduced new 400G DR4 and 2x400G DR4 OSFP modules, which enable direct NIC-to-switch connections and support AI workloads that demand low latency. By developing and adopting these cutting-edge technologies, Meta is creating a more efficient, scalable, and flexible AI infrastructure that can support the growing demands of its AI workloads.
