Microsoft unveils 'world's most powerful AI datacenter'

"Microsoft is opening the largest AI datacenter ever built in Wisconsin. The 315-hectare Fairwater complex serves as the foundation for a global network of purpose-built AI factories that collectively function as a massive AI supercomputer. Fairwater runs on Nvidia GB200 servers in a configuration that includes 72 Blackwell GPUs per rack. Each rack generates 1.8 terabytes of GPU-to-GPU bandwidth and gives each GPU access to 14 terabytes of shared memory. This architecture processes 865,000 tokens per second-the highest processing level of any cloud platform."

"The networking infrastructure combines NVLink and NVSwitch technology at the rack level with 800 Gbps InfiniBand and Ethernet fabrics between racks. The design eliminates congestion by providing each GPU with full-line rate communication with all other GPUs. To minimize physical latency, Microsoft is implementing a dual-layer datacenter design. Racks connect not only to adjacent systems, but also to hardware on other floors. This approach significantly reduces the physical distance between components."

"The complex spans 315 hectares and houses three buildings with a combined floor area of 1.2 million square meters. Its construction required 75 kilometers of foundation piles, 12 million kilograms of steel, and 193 kilometers of underground cables. This facility is fundamentally different from traditional cloud datacenters. Whereas ordinary datacenters run much smaller, independent workloads such as websites and email, Fairwater functions as one giant AI supercomputer."

Fairwater occupies 315 hectares and comprises three buildings with 1.2 million square meters of floor space, supported by extensive foundation piles, steel, and underground cabling. The site runs Nvidia GB200 servers with 72 Blackwell GPUs per rack; each rack generates 1.8 terabytes of GPU-to-GPU bandwidth and gives each GPU access to 14 terabytes of shared memory. The architecture processes 865,000 tokens per second. Networking combines NVLink and NVSwitch within racks with 800 Gbps InfiniBand and Ethernet between racks to eliminate congestion and provide full-line rate GPU communication. A dual-layer rack design connects racks to adjacent systems and other floors to reduce physical latency.

Read at Techzine Global