Dell PowerSwitch SN5610 800GbE Switch

The Dell PowerSwitch SN5610 is a 2U spine and super-spine switch with 64 OSFP ports of 800GbE plus two 25GbE SFP28 ports, built on the NVIDIA Spectrum-4 ASIC, delivering 51.2 Tbps of bidirectional switching capacity, 33.3 Bpps of forwarding and a 160MB fully shared packet buffer.

It is one of two 800GbE models in the SN5000 series and the one specified for standard AC facilities. It is also a Spectrum-X platform switch, which is a specific technical claim rather than a marketing tier, and the reason to choose it over the otherwise-similar SN5400.

What Spectrum-X membership actually means here

NVIDIA's own sentence is precise, and the parenthesis in it is load-bearing: the Spectrum-X platform "integrates Spectrum 5000 series Ethernet switches (excluding SN5400) with NVIDIA SuperNICs, standing out as the world's first Ethernet fabric built specifically for AI."

The SN5610 is in. The SN5400 is not. Both run Spectrum-4 silicon and both carry the same 160MB buffer, so on a spec sheet comparison the difference reads as bandwidth alone. It is not. Spectrum-X is a switch-plus-NIC system - the switches are pretested and prevalidated with BlueField SuperNIC cards and DPUs, ConnectX SuperNICs and NICs, and LinkX interconnects - and NVIDIA attributes a 1.6x improvement in generative-AI network performance over conventional Ethernet fabrics to that end-to-end codesign, along with performance isolation across thousands of simultaneous AI jobs, multi-tenancy and zero-trust security.

If the design assumes Spectrum-X behaviour, the switch has to be a Spectrum-X switch. If the fabric is a conventional Ethernet build, the exclusion costs nothing and the SN5400 is the cheaper 400GbE option. This is the single most useful distinction in the SN5000 series and it is easy to miss because both models sit on the same datasheet.

Sixty-four 800GbE ports and the breakout map

Each of the 64 cages is an OSFP running 8 lanes of 100G PAM4, and each supports 100, 200, 400 or 800GbE. Dell publishes the reachable densities:

• 64 x 800GbE native OSFP
• 128 x 400GbE
• 256 x 200GbE, and the same figure at 100 and 50GbE
• 128 x 40GbE
• 256 + 2 x 25GbE and 10GbE - the +2 being the SFP28 ports

Two hundred and fifty-six ports of 100GbE from two rack units is the number that reframes the switch. That is not an incremental improvement on a 100GbE aggregation switch; it is a different order of consolidation, and it is why NVIDIA positions the SN5000 series as suitable for replacing modular chassis switches. By combining Spectrum-4's low latency with a radix this large, thousands of hosts fit inside a two-tier network - which keeps port-to-port latency minimal because there is no third tier to cross.

The two SFP28 ports are not a fabric tier. They are management-plane and low-speed connectivity, and quantity conversations should treat them as such.

What Spectrum-4 does for AI traffic

Spectrum-4 is the fifth-generation NVIDIA Ethernet switch ASIC. Three of its properties matter for a GPU pod.

The buffer. 160MB, fully shared and monolithic across all ports, supporting cut-through traffic from every port without compromising scale or features. AI training traffic is synchronised by construction - every node completes its step at roughly the same moment and transmits at roughly the same moment - so incast is the steady state, not an exception. A monolithic buffer absorbs it where a partitioned one would not.

RoCE with Spectrum-X extensions. RDMA over converged Ethernet with extensions specifically for AI cloud servers built on Ethernet, plus end-to-end nanosecond-level timing precision from switch to host. Lossless behaviour is what allows a collective operation to complete without a retransmission stalling every participant.

Scale that moves with the design. 512,000 entries dynamically shared across MAC, ARP, IPv4 routes, ACLs, ECMP and VXLAN tunnels rather than fixed at manufacture, plus 256-way ECMP, EVPN multi-homing, multi-chassis LAG for active/active Layer 2 multipathing, and hardware-assisted in-service software upgrades so the fabric can be patched without a maintenance window that stops a training run.

Visibility is the usual Spectrum stack - What Just Happened event-triggered telemetry, hardware-accelerated queue-depth histograms powered by NetQ at sub-microsecond granularity, inband network telemetry, streaming telemetry, and 512,000 on-chip flow counters. Security is a built-in hardware root of trust authenticating everything from BIOS to network operating system; any component or image not signed by NVIDIA is blocked from running.

Power, cooling and physical planning

This is the model with four power supplies. The SN5610 runs 2+2 redundancy where the SN5600D and SN5400 run 1+1 with two. Input is 200-240V AC at 50-60Hz - note the 200V floor, which rules out a low-line installation. Cooling is five hot-swappable fans in N+1 redundancy and the airflow is reverse only; there is no forward-airflow build of this model, so the rack layout has to accommodate it rather than the other way round.

The chassis is 3.39 in high, 17.20 in wide and 31.00 in deep. That 31-inch depth is the constraint most likely to bite: it will not fit a shallow cabinet, and it needs rear clearance for cabling on top of that. The operating envelope is 0 to 40 C, the widest of the three SN5000 models, with non-operating -40 to 70 C, 5-85% relative humidity and altitude to 3,050 m.

NVIDIA describes the SN5610 as introducing a modified hardware specification that optimises system power consumption and improves efficiency relative to the SN5600D. No wattage is published for either model on the family datasheet, and none is invented here - plan the circuit from a measured unit. The datasheet also publishes no weight and no latency figure for any SN5000 model, so freight and latency budgets both need a physical unit or a Dell quote rather than a spec-sheet number.

Software

Two supported network operating systems: NVIDIA Cumulus Linux and Enterprise SONiC Distribution by Dell Technologies, on an ONIE platform that will install alternatives. NVIDIA Air builds a digital twin of the switch and every cable in the topology so automation, policy, monitoring, interoperability and upgrade procedures can be validated before the hardware arrives - which on 800GbE lead times is not a luxury. On a secondary-market unit the installed image tells you nothing about licensing; confirm the NOS before purchase.

Where it sits in the range

The SN5600D is the same 800GbE switch fed from a DC bus bar instead of AC, aimed at DGX SuperPOD racks. The SN5400 is the 400GbE member of the family and the one outside Spectrum-X. The SN4700 is the previous Spectrum-3 generation at 32 x 400GbE in 1RU. On the Dell-branded side the Z9864F-ON is the comparable 800GbE PowerSwitch. If the compute fabric is InfiniBand rather than Ethernet, the equivalent tier is the Quantum-X800 Q3400-RA; the Dell AI fabric switch guide compares the two paths. Whichever fabric you choose, the out-of-band network is Ethernet, and that is the SN2201. Spec sheets for all of them sit on our AI cluster networking lineup.

Every unit is professionally reconditioned, fully tested and shipped with warranty. This model is quote-only. An 800GbE spine switch of this generation barely circulates on the secondary market, so tell us the port map, optics plan and quantity and a specialist will come back to you on sourcing and availability.

Specifications