NVIDIA Spectrum-X SN5600 Ethernet Switch

The NVIDIA Spectrum-X SN5600 is a high-density 800GbE Ethernet switch engineered from the ground up to meet the extreme bandwidth and latency demands of large-scale AI and machine learning training clusters built around NVIDIA Blackwell-architecture GPUs. Powered by the NVIDIA Spectrum-4 ASIC, the SN5600 delivers up to 51.2 Tbps of non-blocking switching capacity across 64 ports of 800GbE or 128 ports of 400GbE, providing the density necessary to build fat-tree and dragonfly topologies at scale without oversubscription. The switch natively supports Spectrum-X Enhanced Ethernet, NVIDIA's purpose-built AI networking technology that combines hardware-based adaptive routing and NVIDIA DCQCN-X congestion control to keep every GPU continuously fed with data, dramatically reducing idle GPU cycles caused by network-induced stalls.

At the heart of the SN5600's AI optimization is its ability to perform per-packet adaptive routing decisions entirely in hardware, dynamically load-balancing traffic across all available equal-cost paths in real time without requiring flow pinning. This eliminates the head-of-line blocking and incast congestion that plague conventional Ethernet switches in all-to-all collective communication patterns—such as AllReduce—that are fundamental to distributed deep learning. The integrated telemetry engine provides nanosecond-resolution per-port and per-flow visibility, feeding NVIDIA UFM (Unified Fabric Manager) with the granular data required for proactive congestion avoidance and rapid fault isolation across multi-thousand-GPU fabrics.

The SN5600 is designed for seamless integration into NVIDIA DGX SuperPOD and NVIDIA BasePOD reference architectures and is managed through NVIDIA UFM Software, NVIDIA Air, and standard industry frameworks including OpenConfig, gRPC/gNMI, and SONIC-based NOS options. Its 2U form factor, front-to-rear airflow, and hot-swappable power supplies and fans align with hyperscale and enterprise data center operational requirements. By combining silicon-level AI awareness with open management interfaces, the SN5600 enables operators to maximize GPU utilization—translating directly into faster model training completion times and higher return on AI infrastructure investment.

Ideal for

• Large-scale NVIDIA Blackwell GPU cluster interconnect fabric for distributed LLM pre-training workloads requiring maximum all-to-all collective communication throughput
• NVIDIA DGX SuperPOD and BasePOD leaf/spine switching layer providing non-blocking 800GbE connectivity between DGX H200 and DGX B200 nodes
• Hyperscale AI cloud provider multi-rack GPU pod fabrics leveraging adaptive routing to sustain near-theoretical bisection bandwidth under heavy AllReduce traffic
• Enterprise AI data center spine switches aggregating multiple 400GbE downlinks from GPU servers into high-capacity 800GbE uplinks for east-west AI traffic
• High-performance computing clusters running mixed AI inference and simulation workloads that demand ultra-low latency and per-flow traffic telemetry for SLA enforcement
• Research institution AI supercomputer fabrics requiring open NOS flexibility via SONiC alongside NVIDIA UFM for end-to-end fabric observability and automated congestion remediation

Technical specifications

Available from Omnixon Global. Submit an RFQ and our team will confirm configuration and availability for your order.