DriveNets says its AI Fabric has been deployed to connect two WhiteFiber H200 GPU cluster data centers 52 miles apart into what the companies describe as a single logical GPU supercluster. DriveNets reports the setup was validated at 111.2 Tbps of bandwidth with 0.9 ms of guaranteed latency.
The deployment is tied to WhiteFiber’s “Project Redwood,” which links the two geographically separated GPU clusters and uses DriveNets AI Fabric as the network between sites. DriveNets characterizes this as a commercial deployment of long-distance “scale-across” AI networking, rather than a lab proof-of-concept.
Stretching AI training clusters across distance isn’t the same problem as traditional data center interconnect. AI training traffic tends to arrive in synchronized bursts and in very large flows, which can expose buffering and congestion-management limits on inter-site links. If that traffic turns into congestion, latency spikes and packet loss can stall training jobs, leaving GPUs on both sides underutilized. The hard part here is keeping behavior predictable across a long-distance link while maintaining high utilization.
DriveNets points to its 9300F, 5300R, and 5301R switches and its Fabric Scheduled Ethernet (FSE) technology as the basis for extending an AI fabric beyond a single facility. The company lists cell-based load balancing, end-to-end Virtual Output Queuing (VOQ), and a deep-buffer interconnect as mechanisms intended to absorb bursts before they cause congestion.
“Power availability can be a major limit to AI infrastructure growth, but with this proven deployment, it no longer has to be,” said Ido Susan, co-founder and CEO of DriveNets. “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.”
“DriveNets’ AI Fabric was critical to proving that Project Redwood could deliver the performance and reliability of a single-site cluster across two locations,” said Sam Tabar, CEO of WhiteFiber.
DriveNets says the validation compared performance between GPU racks within a single site versus performance between GPU racks split across the two sites, with one rack at the primary site and one at the remote site. The company points to a white paper for additional methodology and results.
Source: DriveNets
