Dell PowerSwitch S4348T-ON 10GbE Switch

The Dell PowerSwitch S4348T-ON is a 1RU top-of-rack switch with 48 ports of 10GBASE-T copper on RJ45 and six multi-rate 100GbE QSFP28 uplinks, forwarding at 1001.7 Mpps across a 1.08 Tb/s non-blocking fabric with a 32MB buffer and hardware VXLAN gateway support, running Enterprise SONiC Distribution by Dell Technologies.

It is the copper member of Dell's current 10GbE top-of-rack pair, and it exists for one reason that has nothing to do with performance: the cable is already there.

The case for 10GBASE-T is a cabling case

A 10GBASE-T port takes the structured copper already installed in the rack and the building - Cat6a in a modern install - and needs no transceiver, no fibre patching and no optics inventory. For a rack of servers with 10GBASE-T LOM ports, that means a patch cable and nothing else. Against a fibre switch of the same port count, the saving on optics alone across 48 ports is the largest single line in the comparison, and it repeats every time a port is added.

It is also the switch that survives contact with reality in mixed environments. RJ45 negotiates down: a 10GBASE-T port will happily run 1GbE and, in most implementations, 5G and 2.5G, so one switch can serve a rack where not every server has been refreshed. A fibre switch cannot do that without a different transceiver in every port.

The trade is real and Dell publishes it, which is unusual and worth using.

What copper costs you, in Dell's own numbers

The S4348-ON specification puts this model and its SFP+ twin in one table, so the copper penalty is measured rather than argued.

Latency: sub-3.5 microseconds on this model against sub-650 nanoseconds on the S4348F-ON - roughly five times higher. That is inherent to 10GBASE-T: the physical layer does substantial signal processing to push 10 gigabit down twisted pair, and that processing takes time. For nearly all enterprise workloads a few microseconds at the switch is irrelevant against the milliseconds spent everywhere else. For high-frequency trading, tightly-coupled HPC or anything measuring end-to-end latency in single-digit microseconds, it disqualifies the switch. Know which of those you are.

Power: 222W typical operating against 119W - nearly double, for the same fabric and the same uplinks. Again this is the PHY: driving 10 gigabit over copper takes real energy per port, and it does not go away when the link is idle in the way an unlit optic does. Over a fleet and a few years that difference is a line item, and it is worth putting into the comparison alongside the optics saving that pushed you towards copper in the first place.

Weight: 19.84 lbs against 20.28 lbs with two power supplies and five fan trays - effectively identical, and mentioned only because it confirms these are the same chassis with a different front panel.

Everything else is the same: 1.08 Tb/s fabric, 1001.7 Mpps, 32MB buffer with integrated SmartBuffer, 16GB of CPU memory, six multi-rate 100GbE uplinks, hardware VXLAN gateway, IEEE 1588v2 in hardware, and the same scale numbers.

What changed from the S4148T-ON

The previous-generation equivalent is the S4148T-ON, and the differences are in the same places as on the fibre side.

Uplinks: the S4148T-ON carries four 100GbE QSFP28 plus two 40GbE QSFP+ - 480 Gbps across two port types. This switch carries six 100GbE QSFP28 and no QSFP+ at all: 600 Gbps, one port type, one optics inventory. 40GbE is still available, just as a speed on a multi-rate 100G port rather than as a dedicated one.

Buffer: 12MB becomes 32MB. On a top-of-rack switch that is the number that decides how the rack behaves under incast, when many servers answer one request at once and all burst into the same uplink at the same moment. Storage and distributed-database traffic feel this before anything else does.

Control plane: 4GB of CPU memory becomes 16GB, which is what makes container-based network operating systems, modern telemetry and on-box automation practical.

Software: the S4100 series is an OS10 platform; the S4348-ON series specification refers only to Enterprise SONiC Distribution by Dell Technologies, with ONIE for zero-touch installation. If you are standardising an estate, decide that deliberately rather than inheriting it at delivery.

One caution on headline figures. Dell states this fabric as 1.08 Tb/s half-duplex while the older S4100 sheets state theirs on the full-duplex convention, so a straight numeric comparison between the two documents is not valid - and the same applies to the packets-per-second numbers. Compared like for like the newer switch is the larger fabric. The port arithmetic needs no interpretation: 48 by 10G plus 6 by 100G, all line-rate and non-blocking under full load.

Six multi-rate uplinks

The six QSFP28 ports run 10, 25, 40 or 100 gigabit. On a copper access switch that flexibility does most of its work at the uplink: six 100GbE ports is generous headroom for 480 Gbps of access, enough for redundant uplinks into two spines with ports left over during a failure, and enough to build a small leaf-and-spine directly out of these switches without a separate spine tier. Dell publishes a maximum 10GbE density of 60 for the switch, which is the access ports plus what the uplinks can be subdivided into.

Running some uplinks at 25 or 40 gigabit is the practical migration path for a network that has not finished moving to 100 - it avoids buying an interim switch. On attach, the 100GbE ports are offered with 100GBASE-CR4 passive direct-attach copper; 100G optics over distance, or breakout from a 100G port into four 25G links, are sourcing conversations to have at quotation rather than menu selections.

VXLAN, VRFs and the scale behind them

Dell calls out hardware support for L2 and L3 VXLAN gateway as unique to this series among its 10GbE peers. VXLAN carries Layer 2 segments across a routed underlay, and terminating it in hardware at line rate - for bridging and for routing - lets the top-of-rack switch be the boundary between a virtualised overlay and the bare-metal servers, storage and appliances that are not virtualised, without a software gateway becoming the bottleneck.

The published scale is worth checking against your design: 4K Layer 2 VLANs, 4K Layer 3 interfaces per system, up to 1000 VRFs, 224K MAC addresses in L2 mode and 64K in L3 mode, 32K IPv6 hosts and 32K IPv6 routes rising to 65K in Route Scale mode, 16K ARP entries, 510 RPVST instances, and ACL tables of 768 ingress rules and 512 egress rules for IPv4, IPv6 and MAC alike.

IEEE 1588v2 is supported in hardware, which makes precision time distribution credible for financial, broadcast and industrial-measurement workloads - and which is a genuinely useful pairing with copper access, since those environments are frequently the ones with existing structured cabling they intend to keep.

A note on software: a Dell software entitlement is bound to the unit's service tag and to a Dell account and does not automatically transfer on the secondary market. A refurbished unit may arrive with a working image installed and without an active entitlement or support contract. We do not promise a transferred Dell licence, a software subscription or a support contract with a refurbished switch. Tell us what you intend to run and we will confirm the state of a specific unit before it ships.

Airflow: choose it, and choose it once

The S4348-ON is built in both directions - I/O panel to PSU and PSU to I/O panel - with the power supply and the fan module offered as direction-specific parts in each. Hot-aisle and cold-aisle layouts need opposite directions, and a switch moving air the wrong way through a contained aisle is a thermal problem that gets blamed on everything except the switch.

Two rules. Every airflow part in one chassis must match: five fan trays and two power supplies, all one direction. And match the switch to the aisle it is going into, not to the switch below it, unless both share an orientation. Because the direction is a power-supply and fan swap on one chassis, we sell a single model page and the direction is part of the build - confirm it with your quote request.

Power and cooling

Maximum power consumption is 393W with 222W typical, and Dell publishes a maximum thermal output of 986.14 BTU/h. Power supplies are hot-swappable and redundant, 90-240 VAC at 50/60 Hz, with DC supplies (-40V to -72V DC) available as a custom kit for DC plant. Cooling is five hot-swappable fan trays with two fans each, so a single fan failure is neither an outage nor an emergency.

The 222W typical figure is the one to size a PDU against rather than the 393W ceiling, but it is also the number to carry into any comparison against a fibre switch - copper access costs about 100W of continuous draw at this port count, and that is the honest counterweight to the optics saving.

Common questions

Can I stack these?

No. The six 100GbE QSFP28 ports are uplinks. Rack-level redundancy is built from link aggregation - Dell publishes 128 port channels per system with up to 32 members per channel - with multi-chassis behaviour depending on the Enterprise SONiC release you run. Confirm the feature set against your intended software version.

What cabling do the access ports need?

Cat6a for 10GBASE-T at full distance. No transceiver of any kind is required, which is the main practical advantage of this model over the SFP+ version.

Will the ports run slower links?

RJ45 negotiates down, so a 10GBASE-T port serves 1GbE devices as well - useful in a rack that has not been refreshed all at once. Confirm the exact negotiated speeds you need with your quote request, as the supported set is a function of the Enterprise SONiC release.

Should I buy this or the S4348F-ON?

Copper against fibre, with everything else equal. Take this model if your servers have 10GBASE-T ports and the structured cabling exists - you save the optics entirely. Take the S4348F-ON if your servers have SFP+ adapters, if you need the lower latency, or if you want the lower continuous power draw. There is no performance argument between them: same fabric, same buffer, same uplinks.

Is it a drop-in replacement for an S4148T-ON?

It occupies the same position with more uplink capacity, a much larger buffer and a far larger control plane, but it is not a drop-in: the operating system is different and there are no 40GbE QSFP+ ports. If you are extending an existing OS10 estate, the S4148T-ON keeps the fleet uniform.

Is a TAA-compliant build available?

Dell lists both TAA and NON-TAA builds of each airflow direction. Flag the requirement with your quote request so it can be verified against the specific unit.

Where it sits

Availability, condition and how to buy

This model is supplied to order. Airflow direction, power-supply type, fan configuration and uplink attach are confirmed at quotation, so use the quote request on this page and tell us which way the aisle runs.

Every unit is refurbished, professionally reconditioned, powered up, inspected and functionally tested before it ships, and it ships with warranty coverage. If you need a specific Enterprise SONiC state, a TAA build, or a matched set for a leaf tier, say so with your request.

Specifications