AI-Fabric Switching Designed for Scale, Speed, and Simplicity

Montreal, July 2025 — E.C.I. NETWORKS proudly announces the release of its next-generation Ready-to-Plug SONiC AI switches, purpose-built for high-performance AI/ML training and inference workloads. These switches are optimized for RoCEv2 Ethernet fabrics and multi-rail GPU cluster architectures, supporting the most demanding AI deployments—including LLM training and hyperscale inference.

Key Highlights

• Purpose-Built for AI Fabrics
  Supports multi-tier GPU cluster interconnects with lossless RoCEv2, dynamic load balancing, and end-to-end ECMP optimization.
• 800G and 400G High-Radix Designs
  Scalable leaf and spine platforms for 2-tier and 3-tier AI data center topologies.
• SONiC-Ready and Fully Integrated
  Switches ship pre-loaded with enterprise-grade SONiC NOS—ready for plug-and-play deployment across GPU pods and SuperPODs.
• AI-Focused Feature Set
• RoCEv2 with PFC & ECN
• QPN-based ECMP hashing
• Adaptive routing
• Cut-through switching
• L3 VXLAN segmentation

AI Fabric Benefits

• Open Ethernet Architecture – No proprietary lock-in
• SONiC Pre-Integration – Plug-and-play with built-in RoCEv2 tuning
• High-Radix Designs – >51.2Tbps switching capacity for extreme scalability
• Validated Stack – End-to-end optics, cables, and software tested by E.C.I.
• ODM/OEM Hardware – Flexible configurations to match performance and budget needs

Ideal For

• DGX H100-based GPU PODs and SuperPODs
• AI/ML training clusters running LLMs (e.g., GPT-4, GPT-5)
• Telcos, hyperscalers, and research institutions building backend GPU fabrics

Product Portfolio

Licensing & Support

• Enterprise SONiC AI Package
• Available with 3-year or 5-year subscription options with 24/7 support.

Includes PFC, ECN, DLB, QPN-based ECMP, L3 VXLAN (TH5), and more.

For more informtion contact sales@ecin.ca

To learn more about Ready-to-Plug SONiC AI Switches, visit: Home | SONiC