Dell PowerSwitch Z9864F-ON 800GbE Switch

The Dell PowerSwitch Z9864F-ON is a 2U AI fabric switch with 64 ports of 800GbE in OSFP112 form factor plus two 10GbE SFP+ ports, 102.4 Tbps of switching capacity full duplex, 51.2 Tbps non-blocking half duplex, 20.3 Bpps of forwarding capacity, a 165.2MB packet buffer and sub-700 ns latency.

This is the top of Dell's fixed-format Ethernet range and the only 800GbE switch in it. Dell files it under both AI switches and data centre switches, and states its intended scale plainly: up to 8,000 GPU nodes in a single two-tier 400G fabric.

Sixty-four 800GbE ports, and the breakout map

Each of the 64 cages is an OSFP112 running PAM4 signalling, and each supports 100, 200, 400 or 800GbE. Dell publishes the reachable configurations directly:

• 64 x 800GbE native OSFP112 PAM4
• 128 x 400GbE in breakout mode (100GbE PAM4)
• 256 x 200GbE in breakout mode
• 320 x 100GbE in breakout mode
• 256 x 100GbE plus either 32 x 800GbE or 64 x 400GbE as a mixed build

Two of those - the 256 x 200GbE line and the mixed configuration - are footnoted by Dell as provisional breakout support at release to ship. That qualifier is carried here rather than quietly dropped, because a fabric design that assumes it is available is a design that may need revisiting.

The 320-port 100GbE figure is the one that reframes the switch. Three hundred and twenty ports of 100GbE from two rack units is not an incremental improvement on a 100GbE aggregation switch; it is a different order of consolidation.

What makes this an AI fabric switch rather than a fast switch

Raw bandwidth is the least interesting part. Three other things do the work.

Buffer, and how it is divided. The packet buffer is 165.2MB, and Dell quotes the structure as well as the total: 82.6MB per Ingress Traffic Manager, with two ITMs per NPU. 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 the incast is not occasional, it is the normal operating pattern. Buffer is what absorbs it, and knowing how the buffer is partitioned is what lets you predict behaviour instead of hoping.

Lossless transport. Priority flow control, DCBX, enhanced transmission selection and RoCEv2 give RDMA traffic the lossless treatment it needs across a routed fabric. NVMe/TCP and Dell SmartFabric Storage Software extend the same automation to storage services on existing IP fabrics.

Fabric-aware forwarding. Dell documents a set of AI-specific hardware enhancements for this model - versatile hashing, adaptive routing, cognitive routing, enhanced priority-based flow control queue configuration, increased AI observability and Dell SmartFabric Manager support. Every one of those is footnoted by Dell as requiring a future software release for full feature functionality. If any of them is load-bearing in your design, confirm the software state before you commit.

Power and cooling: plan this before you plan the ports

This switch has the most demanding power and thermal profile in the Dell fixed range, and the constraints are interlocking.

Maximum power available is 3000W on AC high line, 1500W on AC low line and 3200W on DC, through an IEC60320 C20 inlet. Maximum utilised power is 2848W, and Dell's typical figure of 1500W is measured at 25 C ambient with 100% load on all ports using DAC - which is the important qualifier, because DAC draws far less than optics. Redundancy has a condition attached: Dell states that 1+1 power redundancy is supported only in high power mode. A low-line installation is not a redundant installation.

The optics power budget is where airflow direction stops being a cabling preference and becomes a design constraint:

• I/O panel to PSU airflow: up to 64 ports of 18W optics.
• PSU to I/O panel airflow at 35 C: up to 32 ports of 17W optics - half the chassis.
• PSU to I/O panel airflow at 45 C: up to 64 ports of DAC cable, and no optics support at all.

Read that last line twice. In reverse airflow at the top of the thermal envelope, this switch is a DAC-only switch. If your rack layout requires reverse airflow and your links require optics, the ambient temperature is now part of your fabric design. Dell sells the chassis as four airflow kits - AC or DC, in either direction, each with two power supplies and four fans - so the decision is made at order time, not in the rack.

Cooling is four hot-swappable fan modules at 3+1 redundancy, each module carrying one dual-stacked fan.

SONiC, and why that is different here

The "-ON" suffix is Dell's open networking designation, but this model is not interchangeable with its siblings on the software question. Dell's ordering table for the Z9864F-ON names one network operating system: Enterprise SONiC Distribution by Dell Technologies. SmartFabric OS10 is not listed. The ONIE bootloader will install alternative network operating systems "when applicable" - Dell's own hedge, and worth taking literally.

If you are standardised on OS10 elsewhere in the estate, that is a real planning item rather than a footnote. Confirm the intended NOS before purchase.

Underneath sits a 4-core Intel Xeon D-1714 control plane at 2.3GHz with turbo to 3.4GHz, 32GB of DDR4 ECC, an NPU pipeline that is programmable in NPL, and a monolithic 5nm die package. Security is Hardware Root of Trust with Secure Boot and TPM 2.0. Timing is comprehensive: a SyncE card with a Stratum 3E class OCXO, and PTP/1588 delivering Class C phase performance or better.

Physical planning and what is in the box

The chassis is 3.43 in high, 17.26 in wide and 24.80 in deep, weighing 55.8 lbs with power supplies and fan trays installed or 44.1 lbs without. That is a genuine two-person lift. Operating range is 32 to 113 F (0 to 45 C) at 8 to 90% relative humidity.

Usefully, and unlike the rest of the family, each unit ships with console cables and an Accelerated Rack Mount rail kit included - so there is no separate rail or console cable to source.

Where it sits in Dell's range

The Z9664F-ON is the 400GbE step below, with 64 x 400GbE and a 51.2 Tbps full-duplex fabric in the same 2RU footprint - and the duplex qualifier is the distinction that matters between the two, not the shared 51.2 Tbps number. The Z9432F-ON is the 1RU 400GbE option, the Z9332F-ON is the first-generation 400GbE model, and the Z9264F-ON and Z9100-ON hold the 100GbE tier.

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

Specifications