Microsoft Brings Fairwater GB200 Supercomputer Online, Delivering 10x Performance Over Top500 Systems

Technology firm Microsoft introduced a brand new wave of purpose-built datacenters and infrastructure investments worldwide to speed up the adoption of superior AI workloads and cloud companies. The firm unveiled its newest US AI datacenter, described as the biggest and most superior AI facility it has constructed to this point. Alongside the Fairwater datacenter in Wisconsin, a number of further Fairwater services are underneath growth throughout totally different US areas.

Internationally, Microsoft revealed plans in Narvik, Norway, in collaboration with nScale and Aker JV, to determine a hyperscale AI datacenter. In the United Kingdom, the corporate introduced a partnership with nScale to construct what is ready to be the nation’s largest supercomputer, primarily based in Loughton, to help home companies.

These AI datacenters signify multibillion-dollar capital initiatives outfitted with lots of of 1000’s of high-performance AI chips. They will combine seamlessly with the Microsoft Cloud community, which spans greater than 400 datacenters throughout 70 world areas. By interlinking these new AI services right into a distributed system, Microsoft goals to multiply effectivity and computing capability, additional increasing entry to AI companies on a world scale.

AI datacenters are specialised services designed for coaching and deploying large-scale AI fashions and purposes. Microsoft’s AI datacenters help workloads similar to OpenAI fashions, Microsoft AI, Copilot options, and different superior AI methods. The newly accomplished Fairwater datacenter in Wisconsin illustrates the dimensions of those initiatives, occupying 315 acres and comprising three massive buildings with a complete of 1.2 million sq. toes of area. Its development required intensive infrastructure, together with 46.6 miles of basis piles, 26.5 million kilos of structural metal, 120 miles of underground medium-voltage cable, and 72.6 miles of mechanical piping.

Unlike conventional cloud datacenters, which usually run quite a few smaller workloads like web sites, e-mail, or enterprise purposes, Microsoft’s new facility is engineered to function as a large AI supercomputer. Featuring a flat networking structure interconnecting lots of of 1000’s of NVIDIA GPUs, it’s anticipated to ship ten occasions the efficiency of right this moment’s quickest supercomputer, enabling AI coaching and inference at an unprecedented scale.

Microsoft’s Wisconsin Datacenter To Serve As Core Of Global AI Supercomputer Network

Microsoft highlighted the function of purpose-built infrastructure in effectively powering superior AI fashions on the trillion-parameter scale. At the core of its AI datacenters are GPU accelerators built-in with CPUs, reminiscence, and storage, organized into racks and related via low-latency networking. While this setup seems as unbiased servers, at scale it operates as a single supercomputer able to coaching large fashions in parallel.

The Wisconsin facility runs a unified cluster of NVIDIA GB200 servers with thousands and thousands of compute cores and exabytes of storage, enabling processing speeds of as much as 865,000 tokens per second—presently the best cloud throughput out there. Each rack accommodates 72 Blackwell GPUs related via NVLink, pooling 14 terabytes of reminiscence per GPU. Future datacenters in Norway and the UK will undertake comparable clusters utilizing NVIDIA’s next-generation GB300 chips with expanded reminiscence capability.

In order to take care of effectivity at supercomputing scale, Microsoft deploys superior networking that permits GPUs inside and throughout racks to speak at terabytes per second with out congestion. Multiple pods of racks are then linked to perform as one global-scale supercomputer. The Wisconsin datacenter makes use of a two-story configuration to reduce latency from bodily distance, additional bettering connectivity.

Through this layered design, co-engineered with business companions, Microsoft Azure has established what it describes because the world’s strongest purpose-built AI supercomputer, designed to help frontier fashions and large-scale AI workloads.

Microsoft’s new AI datacenters type a part of a globally interconnected Azure community, linked via a large space system designed to perform as a single, large-scale AI supercomputer. By distributing compute, storage, and networking sources throughout a number of areas, the infrastructure delivers higher scalability, resilience, and suppleness than a single facility may present.

This shift required an entire redesign of the cloud stack, aligning {hardware} and software program throughout silicon, servers, networks, and datacenters right into a unified, purpose-built system. The Wisconsin datacenter performs a central function on this imaginative and prescient, combining superior expertise and funding whereas contributing to regional growth. Integrated with different services worldwide, it helps a brand new part of safe and adaptive cloud-based AI.

The put up Microsoft Brings Fairwater GB200 Supercomputer Online, Delivering 10x Performance Over Top500 Systems appeared first on Metaverse Post.