NVIDIA Spectrum-3 SN4600 64x100G/32x200G Ethernet Switch

The NVIDIA Spectrum-3 SN4600 is a high-radix 400G-class spine switch built on the third-generation Spectrum ASIC, delivering 64 ports of 100GbE or 32 ports of 200GbE in a compact 2U form factor. The switch provides 12.8 Tbps of aggregate switching capacity with deep packet buffers, hardware-based telemetry, and full support for RoCEv2 (RDMA over Converged Ethernet), making it an ideal fabric element for latency-sensitive, bandwidth-intensive AI training and high-performance computing environments. The Spectrum-3 ASIC implements cut-through forwarding with sub-microsecond latency and supports adaptive routing to dynamically balance traffic across available paths without requiring manual flow steering.

The SN4600 operates under NVIDIA's Onyx network operating system, which provides a rich CLI experience alongside full OpenConfig and YANG model support for programmatic management. The switch integrates natively with NVIDIA Air, UFM (Unified Fabric Manager), and the broader DOCA software ecosystem, enabling centralized congestion control, fabric telemetry aggregation, and automated network optimization across large-scale GPU clusters. ECMP and weighted ECMP are supported at line rate across all port configurations, and the platform includes hardware support for VXLAN bridging and routing to extend overlay network capabilities into the data plane.

Designed for hyperscale AI infrastructure and enterprise HPC spine tiers, the SN4600 supports spine-leaf topologies where its 64-port density at 100G allows broad connectivity to leaf-layer switches with high oversubscription freedom. The platform's support for NVIDIA Scalable Hierarchical Aggregation and Reduction Protocol (SHARP) in-network computing offloads collective operations—such as AllReduce—directly into the network fabric, reducing GPU idle time during distributed training workloads. Redundant hot-swappable power supplies and fans, with front-to-back or back-to-front airflow options, support deployment in standard data center rack configurations.

Ideal for

• Spine-layer switching in large-scale distributed AI/ML training clusters connecting GPU servers over RoCEv2 with hardware congestion management
• High-performance computing fabric interconnect for MPI-based parallel workloads requiring consistent low-latency and high-bandwidth paths between compute nodes
• Data center spine aggregation in spine-leaf architectures requiring high port density at 100G or 200G to aggregate multiple leaf switches
• In-network computing environments leveraging NVIDIA SHARP to offload collective communication operations and reduce GPU synchronization overhead
• Telemetry-rich network fabric supporting streaming per-flow and per-port statistics collection for real-time congestion visibility and capacity planning
• Scalable VXLAN overlay backbone for multi-tenant enterprise data centers requiring hardware-accelerated VTEP functionality at spine scale

Technical specifications

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