Dell PowerSwitch Z9664F-ON 400GbE Switch

The Dell PowerSwitch Z9664F-ON is a 2RU aggregation switch with 64 ports of 400GbE in QSFP56-DD form factor plus two 10GbE SFP+ ports, a 51.2 Tbps full-duplex switch fabric, 10.2 Bpps of forwarding capacity, a 113.66MB packet buffer and sub-850 ns latency.

Dell lists it under both AI switches and data centre switches, and the reason is in the numbers: 64 x 400GbE is the port count at which a two-tier fabric stops being an aggregation layer and starts being a spine for something very large.

Sixty-four 400GbE ports, and what they add up to

The headline configuration is 64 ports of 400GbE. With breakout, the same 64 cages reach up to 256 ports of 100GbE, or a mix across 10, 25, 40, 50, 100 and 200GbE with a 256-port ceiling. Every 400GbE port is multi-rate across all of those speeds.

Two of those numbers are worth sitting with. 256 ports of 100GbE in two rack units is a whole aggregation tier in one chassis. And 64 x 400GbE against the 32 x 400GbE of the Z9432F-ON is not just twice the ports - it is twice the fabric, 51.2 Tbps against 25.6 Tbps, which is what lets the port count be used rather than oversubscribed.

Two 10GbE SFP+ ports sit alongside for management uplinks and out-of-band paths that should not consume a 400GbE cage.

The buffer, and why it is quoted per traffic manager

The packet buffer is 113.66MB - and Dell quotes it in an unusually specific way: 56.83MB per Ingress Traffic Manager, of which there are two. That split matters. It means the buffer is not one shared pool that any port can draw on without limit; it is divided across the ingress domains, and a congestion event is served by the buffer belonging to its own half of the chip.

For anyone sizing this switch for incast - many senders converging on one receiver at line rate, which is exactly what distributed training and storage replication produce - the per-ITM figure is the number to design against, not the total. Most listings quote the total and stop. Dell states both, so this page states both.

Alongside the buffer, the switch carries the convergence features that make lossless Ethernet work: Data Center Bridging with priority flow control, DCBX and enhanced transmission selection, plus iSCSI and FCoE support and NVMe/TCP with Dell SmartFabric Storage Software for automating storage services across existing IP fabrics.

Why Dell calls this an AI switch

The AI classification is Dell's own, applied to this model in Dell's own catalogue, and it is earned by three things rather than by marketing. The fabric is large enough that a two-tier design does not bottleneck at the spine. The buffer is large enough to absorb synchronised bursts. And the fabric automation - SmartFabric Services for autonomous deployment, expansion and lifecycle management - is what makes a fabric of this size operable by a team rather than by a heroics rota.

What it is not is a special-purpose box. The Z9664F-ON is a general 400GbE aggregation switch that happens to be well suited to AI fabrics, which is a better thing to buy second-hand than a purpose-built appliance, because its resale and redeployment options stay open.

Airflow, optics power and the constraint that links them

Both airflow directions are orderable: I/O panel to PSU, or PSU to I/O panel. The direction you want depends on which aisle the port face looks into - ports in the cold aisle means I/O to PSU, ports in the hot aisle means PSU to I/O.

On this model that choice also constrains your optics. Dell is explicit that units configured PSU-to-I/O face tighter power restrictions on the cables and optics used for the 400GbE ports. In the I/O-to-PSU direction the budget is up to 15W in every QSFP56-DD port, with an allowance for four ports at 20W; in the reverse direction there are restrictions. If you are planning high-power long-reach modules, decide the airflow direction and the optics together rather than separately.

Cooling is four hot-swappable fans at 3+1 redundancy. Power is two hot-swappable supplies at 1+1 - with Dell's own caveat that lowline power is the exception: a 100-127 VAC supply is a non-redundant solution on this switch. Maximum consumption is 2125W. The 500W typical figure Dell publishes is measured at 25 C ambient with 50% load on all ports, and that condition is worth carrying with the number, because a bare "500W typical" against a 2125W maximum would otherwise be misleading.

SONiC, OS10 and buying open networking hardware used

The "-ON" suffix is Dell's open networking designation. The switch supports the open-source Open Network Install Environment for zero-touch installation of Enterprise SONiC Distribution by Dell Technologies, Dell SmartFabric OS10, or an alternative ONIE-compatible network operating system.

What you buy is the hardware: 64 ports of 400GbE silicon, a 4-core Ice Lake-D control plane running at 2.3GHz with turbo to 3.4GHz, 32GB of DDR4 ECC and an NPU pipeline that is programmable in NPL. The operating system is a separate decision made at install time. Expect a used unit to arrive needing an OS install - a switch at the ONIE prompt is in the state Dell intended, not a fault.

Physically it is 3.45 in high, 17.26 in wide and 24.80 in deep, weighing 50.7 lbs with power supplies and fan trays installed - or 40.68 lbs without. That is a two-person lift at the top of a rack. Rack mounting uses Dell's accelerated mounting kits.

Where it sits in Dell's 400GbE line

The Z9432F-ON is the 1RU, 32-port half of this switch and the right choice when 32 x 400GbE is enough. The Z9864F-ON is the step above, moving to 800GbE on OSFP112 cages with a 51.2 Tbps non-blocking half-duplex fabric aimed at large AI training clusters. The Z9332F-ON is the first-generation 400GbE model, and the Z9264F-ON and Z9100-ON hold the 100GbE tier below.

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

Specifications