FS unveiled a new 800G Ethernet-based AI networking solution optimized for next-generation NVIDIA B300 GPU deployments, positioning lossless Ethernet fabrics as a scalable alternative to proprietary interconnect architectures for large AI training environments.
The new platform is built around a 51.2 Tbps RoCEv2 Ethernet architecture using a Spine-Leaf topology designed for high-bandwidth, ultra-low-latency AI clusters. FS said the solution incorporates Priority Flow Control (PFC), Data Center Quantized Congestion Notification (DCQCN), shared buffering, and intelligent load balancing to deliver lossless Ethernet performance for distributed AI training workloads.
The announcement reflects the broader industry shift toward Ethernet-centric AI fabrics as hyperscalers and cloud providers seek more open and operationally familiar networking architectures for scaling GPU clusters. The rise of NVIDIA B300-class systems significantly increases east-west traffic requirements across AI clusters, placing new pressure on congestion management, RDMA efficiency, and storage interconnect performance.
FS said its architecture integrates the unified PicOS® network operating system across the fabric, enabling centralized operational consistency across multiple network tiers and devices. The company also highlighted integration with its AmpCon automation and management software, which provides automated provisioning, unified monitoring, and standardized operational management for AI data center deployments.
The architecture includes a dedicated RDMA storage network intended to accelerate AI training pipelines by reducing storage bottlenecks and improving GPU utilization during large-scale model training.
Key elements of the FS B300 AI Networking Solution include:
• 51.2T high-throughput Ethernet switching
• 800G Spine-Leaf lossless Ethernet fabric
• RoCEv2 support for AI workloads
• PFC and DCQCN congestion management
• Shared buffers and intelligent load balancing
• Dedicated RDMA storage networking
• PicOS unified network operating system
• AmpCon automation and monitoring platform
While InfiniBand remains dominant in many frontier AI training environments, Ethernet-based AI fabrics continue to gain momentum as vendors push higher-speed switching, improved congestion control, and operational simplification. The emergence of 51.2T switching platforms marks another major transition point for AI infrastructure, enabling denser GPU cluster scaling while reducing oversubscription challenges inside large AI fabrics.
FS, founded in 2009 and headquartered in Delaware, has historically focused on high-speed networking and optical infrastructure for enterprise, telecom, ISP, and cloud environments. The company has increasingly expanded into AI infrastructure networking as demand accelerates for cost-effective, high-capacity Ethernet fabrics capable of supporting hyperscale AI deployments.
| Profile: FS B300 AI Networking Solution | Details |
|---|---|
| Company | FS |
| Headquarters | New Castle, Delaware |
| Launch Focus | 800G Ethernet AI networking for NVIDIA B300 GPU clusters |
| Fabric Architecture | 51.2T RoCEv2 lossless Ethernet Spine-Leaf fabric |
| Primary Use Case | Large-scale AI training clusters |
| Network Speed | 800G Ethernet |
| Congestion Management | PFC, DCQCN, intelligent load balancing |
| Storage Connectivity | Dedicated RDMA storage network |
| Network OS | PicOS unified network operating system |
| Automation Platform | AmpCon data center network management |
| Target Customers | AI cloud providers, enterprises, hyperscalers, telecom operators |
| Industry Trend | Migration toward open Ethernet AI fabrics for hyperscale GPU clusters |
