Dell PowerSwitch Z9664F-ON 64-port 400GbE Switch

The Dell PowerSwitch Z9664F-ON is a 64-port 400GbE fixed-configuration spine switch built on the Broadcom Tomahawk 4 (BCM56990) ASIC, delivering 51.2 Tbps of non-blocking switching capacity across a 1RU form factor. Designed natively for open networking, it runs Dell's Enterprise SONiC distribution or is available as a bare-metal platform supporting third-party NOS deployments, giving data-center architects full programmatic control through open APIs and standards-based tooling.

Targeted at AI/ML training fabric interconnects and hyperscale-style leaf-spine architectures, the Z9664F-ON supports 64 x QSFP-DD ports capable of operating at 400GbE, with breakout flexibility down to 100GbE and 40GbE per port, enabling phased network migrations without chassis replacement. The switch supports ECMP, VXLAN EVPN, SR-IOV-aware forwarding, and deep buffer management features relevant to lossless RoCEv2 and RDMA workloads commonly encountered in GPU cluster interconnects.

The Z9664F-ON integrates with Dell's Open Automation framework and is compatible with standard streaming telemetry via gNMI/gRPC, enabling real-time visibility into fabric health, queue depths, and traffic analytics. Redundant hot-swappable power supplies and fan modules support high-availability deployments, while the platform's power efficiency and thermal design are optimized for high-density spine deployments where rack power budgets are constrained.

Ideal for

• AI/ML GPU cluster spine interconnect providing high-bandwidth, low-latency lossless fabric for distributed training workloads over RoCEv2
• Hyperscale-style leaf-spine data-center core layer delivering 51.2 Tbps non-blocking capacity between top-of-rack leaf switches
• Data-center fabric modernization, leveraging 400GbE port breakout to interoperate with existing 100GbE infrastructure during phased upgrades
• Multi-tenant cloud infrastructure requiring VXLAN EVPN-based network virtualization and isolated tenant overlay segments at scale
• High-performance computing (HPC) interconnect fabrics where deterministic latency and deep buffer management are critical for MPI workloads
• Open networking deployments where operators require bare-metal flexibility to run validated third-party or custom NOS images with full programmatic API control

Technical specifications

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