Dell PowerSwitch SN4700 400GbE Switch

The Dell PowerSwitch SN4700 is a 1U spine and super-spine switch with 32 QSFP-DD ports of 400GbE, built on the NVIDIA Spectrum-3 ASIC, delivering 12.8 Tbps of bidirectional switching capacity, 8.4 Bpps of forwarding and 620 ns port-to-port latency from a 64MB fully shared packet buffer.

Dell describes it as the fourth generation of NVIDIA Ethernet switches. In a fabric it is the layer above the leaf: 32 ports of 400GbE is 128 ports of 100GbE once broken out, which is enough spine capacity for a mid-sized accelerated pod in a single rack unit.

QSFP-DD, not OSFP - check this before you order optics

The connector on the SN4700 is QSFP-DD. Its larger siblings in the SN5610 and SN5600D use OSFP, and the two cages are not interchangeable. Ordering OSFP optics or OSFP direct-attach cables for an SN4700 produces a box of parts that will not seat.

The distinction is worth stating plainly because the 400G and 800G generations are usually discussed together as "OSFP-era" hardware, and they are not. The QSFP-DD cage on this switch runs 8 lanes of 50G PAM4, which is what produces 400Gb/s per port and 12.8 Tb/s across the chassis. The OSFP cage on the 800G models runs 8 lanes of 100G PAM4. Same lane count, double the lane rate, different mechanical form factor.

The SN5400 is the other QSFP-DD model in the Dell AI range, so a site standardising on one optic family can pair the SN4700 and SN5400 without splitting its spares inventory.

The port map

Every port is 400GbE natively, and port speeds span 1 to 400 Gb/s. Dell publishes the reachable densities directly:

• 32 x 400GbE native QSFP-DD
• 64 x 200GbE
• 128 x 100GbE, and the same figure at 50, 25, 10 and 1GbE
• 64 x 40GbE

The 128-port 100GbE line is the one that decides most deployments. A spine that presents 128 ports of 100GbE from 1RU replaces a chassis switch, and the uplink ports allow a variety of blocking ratios so the oversubscription ratio is a design choice rather than a hardware constant.

What makes this an AI fabric switch and not just a fast switch

Three things do the work, and none of them is the headline bandwidth.

The buffer, and how it is shared. 64MB, monolithic, fully shared across all ports, supporting cut-through traffic from every port at once. AI training traffic is synchronised by construction: every node finishes its step at roughly the same instant and transmits at roughly the same instant, so incast is the normal operating pattern rather than an occasional event. A buffer carved into per-port slices runs out on the ports under pressure while sitting idle elsewhere. A fully shared buffer does not.

Lossless transport for RDMA. RoCE support with hardware-accelerated data movement, congestion management and load balancing, aimed explicitly at RDMA and at machine learning applications using GPUDirect. This is the mechanism that lets GPU-to-GPU traffic cross a routed Ethernet fabric without the retransmissions that would otherwise stall a collective operation.

Visibility that survives contact with a real fabric. What Just Happened (WJH) telemetry reports the when, what, who, where and why of a drop rather than leaving an operator to infer it from counters. Hardware-accelerated histograms summarise queue depths at sub-microsecond granularity. The hardware is Inband Network Telemetry-ready, streaming telemetry feeds NetQ, and there are 512,000 on-chip flow counters. On a fabric where one slow link degrades an entire training job, being able to prove which link is the reason is the difference between a fix and a week of guessing.

Scale-wise the ASIC carries 512,000 forwarding entries flexibly shared across ACLs, longest prefix match routes, host routes, MAC lists, ECMP and tunnels - the allocation moves with the design rather than being fixed at manufacture. VXLAN scale is 10x the tunnel and tunnel-endpoint count NVIDIA attributes to comparable silicon, single-pass VXLAN routing and bridging avoids the recirculation penalty, and cloud-scale NAT reaches 100,000 sessions. ECMP is 128-way.

Software: an open platform, so confirm the NOS

The SN4700 is an Open Ethernet switch. Dell's framing is that Open Ethernet "breaks the paradigm of traditional switch systems, eliminating vendor lock-in" by letting the operator choose the network operating system rather than accepting the vendor's. In practice that means two supported paths - NVIDIA Cumulus Linux, and Enterprise SONiC Distribution by Dell Technologies - on an ONIE platform that will install alternatives.

On a refurbished unit the installed image is not the licensed image. Confirm which NOS you intend to run, and its licensing, before purchase.

NVIDIA Air is worth knowing about at design time: it builds a digital twin of the SN4700 and of every cable in the topology, so policy compliance, automation, monitoring, interoperability and upgrade procedures can be validated before anything is racked. On a fabric build where the switches are on long lead times, the twin lets the software work start first.

Power, cooling and physical planning

The chassis is 1.72 in high, 16.85 in wide and 22.3 in deep, weighing 25.8 lbs. Power is two hot-swappable supplies in 1+1 redundancy at 50-60Hz, 100-264V AC. Cooling is six hot-swappable fans in N+1 redundancy, and both the supplies and the fans are colour-coded so airflow direction is readable at the rack rather than in a manual.

Reversible airflow is supported. Note the operating envelope is 0 to 35 C - five degrees tighter than the SN2201 management switch that will usually sit in the same rack. Relative humidity 5-85%, altitude to 3,050 m.

Dell's datasheet gives no maximum power draw and no PSU wattage for this model, so plan the circuit from a measured unit rather than from a published figure. Nothing is stated here that Dell does not state.

Where it sits in the range

The SN4700 is the 400GbE spine at the smaller end. The SN5400 doubles the port count to 64 x 400GbE in 2RU on Spectrum-4; the SN5610 and SN5600D move to 64 x 800GbE. On the Dell-branded side, the Z9664F-ON and Z9432F-ON cover similar 400GbE ground with Dell's own SONiC-first PowerSwitch positioning, and the Z9864F-ON is the 800GbE step. If the fabric is InfiniBand rather than Ethernet, the equivalent tier is the Quantum-2 QM9700 - the Dell AI fabric switch guide sets out how that choice is made. Every model named here is listed together on our AI cluster fabric lineup.

Every unit is professionally reconditioned, fully tested and shipped with warranty. This model is quote-only. Tell us the port map, optics plan, airflow direction and quantity you need and a specialist will come back to you on sourcing and availability.

Specifications