English
WhiteFiber unveiled preliminary R&D results for Project Redwood, a distributed GPU supercluster architecture spanning geographically separate data centers. Last week’s field test hit 111.2 Tbps throughput across 83 km dark fiber with a guaranteed 0.9 ms round-trip latency.
WhiteFiber notes the latency sits within 8% of fiber’s theoretical physical propagation limit. The trial only utilized a portion of available fiber spectrum, yet delivered roughly twice the capacity of published full-spectrum competing field trials. The firm plans to activate full spectrum ahead of its Q3 2026 commercial rollout.
WhiteFiber Project Redwood
The joint research was completed with DriveNets and WEKA. DriveNets delivers inter-site Ethernet AI fabric, while WEKA NeuralMesh powers cross-cluster storage and memory layers. WhiteFiber has filed relevant patent applications for the design.
Unifying Multiple Sites Into One Logical GPU Cluster
This architecture enables two data centers to operate as a single logical GPU supercluster, distinct from separate pods linked via standard DCI links. This difference proves critical for AI training and inference workloads, where inter-GPU synchronization, collective communication and shared data access directly limit maximum cluster scale.
Target workloads include deployments constrained by single-site power, cooling, floor space, resilience rules or data sovereignty mandates, alongside edge, telecom and sovereign AI use cases.
DriveNets confirmed the hardware setup: the fabric interconnects two WhiteFiber NVIDIA H200 GPU clusters 52 miles apart, marking the industry’s first production-grade long-distance scale-across AI supercluster validated outside lab environments. Benchmarks comparing intra-rack and cross-site performance are detailed in DriveNets’ white paper.
DriveNets AI Fabric Architecture
Redundant dark fiber paired with DriveNets AI Fabric carries GPU and storage traffic between facilities. Instead of conventional long-haul Ethernet extensions, it adopts scheduled Ethernet built for consistent distributed AI communication.
The cross-site layer leverages DriveNets Fabric Scheduled Ethernet on its 9300F, 5300R and 5301R switches. Cell-based load balancing, end-to-end Virtual Output Queuing and deep-buffer interconnects absorb synchronized AI traffic spikes to eliminate inter-link congestion. The design delivers lossless, predictable cross-site connectivity that sustains high GPU utilization comparable to single-facility deployments.
These mechanisms stabilize collective communication latency over distance while forming a unified compute-storage fabric with native multi-tenant isolation. The fiber link functions as an integrated segment of the AI fabric rather than a standalone transport pipe, minimizing performance and operational gaps caused by physical data center boundaries.
Differentiation From NVIDIA Spectrum-XGS
WhiteFiber’s testbed differs architecturally from NVIDIA Spectrum-XGS, yet aligns with the broader industry push toward scale-across AI infrastructure.
At Hot Chips 2025, NVIDIA CEO Jensen Huang highlighted demand for cross-city, cross-country and intercontinental data center links to build large-scale AI factories. Spectrum-XGS relies on distance-aware congestion control, precise latency tuning and telemetry for steady distributed GPU traffic, with CoreWeave as an early adopter.
WhiteFiber has released more granular metro-scale field metrics than NVIDIA’s public Spectrum-XGS data; its 83 km trial posted 111.2 Tbps and 0.9 ms round-trip latency, though divergent architectures prevent direct side-by-side performance comparison.
Industry-Wide Multi-Site AI Training Advancements
WhiteFiber’s launch coincides with parallel multi-site AI infrastructure development across major cloud and hardware vendors.
Oracle Cloud Infrastructure and NVIDIA demonstrated LLM training across facilities ~1,000 km apart via NeMo Framework and Megatron-Core, achieving over 96% training scalability through hierarchical all-reduce and segmented inter-data-center communication. This work centers on software and system validation instead of metro transport benchmarks.
Google built TPU Multislice to aggregate multiple TPU slices into unified distributed training pools over internal data center optical networks. Gemini training leveraged multiple geographically dispersed sites, with the platform handling workload partitioning and large-scale fault tolerance.
WhiteFiber aims to bring comparable cross-site orchestration to GPU cloud stacks, prioritizing ultra-high-bandwidth, low-latency metro interconnection. It will release further architectural specifications and launch timelines leading up to Q3 2026 commercial availability.
Beijing Qianxing Jietong Technology Co., Ltd.
Sandy Yang/Global Strategy Director
WhatsApp / WeChat: +86 13426366826
Email: yangyd@qianxingdata.com
Website: www.qianxingdata.com/www.storagesserver.com
Business Focus:
ICT Product Distribution/System Integration & Services/Infrastructure Solutions
With 20+ years of IT distribution experience, we partner with leading global brands to deliver reliable products and professional services.
“Using Technology to Build an Intelligent World”Your Trusted ICT Product Service Provider!
Sandy Yang/Global Strategy Director
WhatsApp / WeChat: +86 13426366826
Email: yangyd@qianxingdata.com
Website: www.qianxingdata.com/www.storagesserver.com
Business Focus:
ICT Product Distribution/System Integration & Services/Infrastructure Solutions
With 20+ years of IT distribution experience, we partner with leading global brands to deliver reliable products and professional services.
“Using Technology to Build an Intelligent World”Your Trusted ICT Product Service Provider!