DriveNets unveiled a scale-across networking deployment connecting two WhiteFiber data centers located some 50 miles apart to form a single distributed logical graphic processing unit (GPU) supercluster.
The deployment saw the vendor extend connectivity between two sites, spanning 111.2 Tb/s with sub-millisecond latency.
DriveNets was supporting WhiteFiber's Project Redwood, linking the geographically separated data centers. The latter plans to add additional scale-across ports with a view to achieving 136 Tb/s of bandwidth in this quarter.
DriveNets’ 9300F, 5300R, and 5301R switches, powered by its Fabric Scheduled Ethernet (FSE) technology, were employed to help extend the AI fabric beyond a single data center. Cell-based load balancing, end-to-end virtual output queuing (VOQ), and deep-buffer interconnect were used to help absorb AI traffic bursts before they cause congestion – a major issue for scale-across networking given the distances between nodes.
“Power availability can be a major limit to AI infrastructure growth, but with this proven deployment, it no longer has to be,” DriveNets CEO Ido Susan noted. “Together with WhiteFiber, we have taken scale-across from concept to commercial reality, showing that two remote data centers can perform as a single high-performance supercluster. This is how we expect many next-generation AI infrastructures to be built.”
“DriveNets’ AI Fabric was critical to proving that Project Redwood could deliver the performance and reliability of a single-site cluster across two locations,” WhiteFiber CEO Sam Tabar added. “This milestone shows that geography no longer has to limit the scale of the AI infrastructure we build.”
Scale-across extends computing interconnectivity beyond the cabinet. Its recent proliferation in the infrastructure zeitgeist stems from rising energy and density costs, with data center deployments increasingly marred by constraints amid the desire to build out ever more white space.
As a result, the concept of distributed computing has become more prevalent. It’s far from a new idea, what with projects like Folding@home breaking the exascale barrier using disparate computers.
Existing deployments of scale-across networking largely focus on the hyperscale level, with Amazon Web Services (AWS) Rainer or Microsoft’s Fairwater among the early alternatives to building massive facilities.
The segment is set to grow further, with the likes of Cisco, Nvidia, and Rad among those racing to get ahead in the burgeoning market.
The segment has already drawn some early critics, with Nvidia's SVP of Networking Gilad Shainer telling SDxCentral that deep-buffer, off-the-shelf switches for scale-across being touted by rival vendors are fundamentally wrong for AI workloads.