The Rise of 100G SFP Form Factors

Hyperscale cloud, AI clusters and 5G infrastructure have changed what a switch port is worth. For close to a decade QSFP28 has been the default 100G optical interface, and it is not going away. What is changing is the shape of the connector beside it.

Three newer form factors, SFP-DD, DSFP and SFP112, deliver 100GbE through the mechanical footprint of an SFP rather than a QSFP. They are not replacements. They exist for the deployments where front panel density, electrical simplicity and power per rack unit have become the binding constraint.

The density problem, drawn to scale

The limit on a 1RU switch is not silicon. It is the width of the faceplate. A QSFP28 cage is roughly twice the width of an SFP cage, so a panel that holds 36 QSFP28 ports holds about 72 SFP ports in exactly the same space. The drawing below is a 1RU front panel with both populations at true relative width.

Doubling the server facing port count in the same rack unit is worth most where connection volume drives the design rather than raw switch to switch bandwidth:

• AI clusters where a single GPU server consumes several 100GbE ports for storage, management and east to west traffic
• High density top of rack switching
• 5G fronthaul and edge aggregation
• Hyperscale cloud fabrics
• Storage and accelerator networks

Why one lane can now carry 100G

None of this works without a change in how bits are put on the wire. NRZ encodes one bit per symbol using two voltage levels. PAM4 uses four levels and carries two bits per symbol, which doubles throughput without doubling the baud rate. That is what allows 100GbE to run on one or two electrical lanes instead of the four that QSFP28 requires.

Four connectors, four sets of tradeoffs

Externally these modules look alike. Electrically and mechanically they are not. The clearest way to separate them is by their contact count, since that is what determines how many lanes a connector can carry and whether it can still accept a legacy module.

QSFP28 remains the workhorse. Four parallel 25G NRZ lanes, a mature ecosystem and broad vendor interoperability keep it the default for enterprise, cloud, storage and service provider networks.

SFP-DD takes the traditional SFP connector and adds a second pair of contact rows, going from 20 to 40. It carries 100GbE on two 50G PAM4 lanes while staying mechanically able to accept SFP+ and SFP28 modules, which is why cloud providers and high density top of rack designs find it attractive.

DSFP reaches similar performance by widening the connector to 22 contacts rather than adding a row. Broad adoption never quite arrived, but it has held its place in telecom infrastructure, particularly 5G fronthaul and midhaul where compact optics and simple manufacturing matter more than backward compatibility.

SFP112 is the cleanest of the four. It carries 100GbE over a single 112G PAM4 lane on the original 20 pin SFP interface, which removes the gearbox, simplifies PCB routing and frees panel space. It is becoming a preferred interface for AI servers, SmartNICs, DPUs and accelerator platforms.

Electrical and mechanical comparison of the four 100GbE interfaces. Scroll the table sideways to see every column.

Where each one has actually landed

Adoption has split along market lines rather than converging on a single winner. Switch vendors have followed their customers, and merchant silicon has followed the switch vendors.

Where each interface has landed, by vendor and market. Scroll the table sideways to see every column.

QSFP28 is still the dominant 100GbE interface in enterprise and cloud networking. SFP-DD and SFP112 are being introduced selectively into next generation AI, cloud and high density platforms rather than displacing QSFP28 across the board.

The cost of packing them closer

Density is not free. An SFP-DD or SFP112 module draws roughly what a QSFP28 draws, between 2 and 3.5 W. Put twice as many of them behind the same faceplate and the power per unit of panel area roughly doubles, while the air available to each module halves.

Platforms designed for this rely on front to back airflow, heat sink and thermal bridge design, active fan control and continuous per module monitoring. Where those are absent, high density optics run hot and their service life suffers.

Connecting new servers to existing switches

Most networks will change over gradually. The practical bridge is a 100G SFP-DD to QSFP28 active optical cable, which lets a next generation server attach to the QSFP28 switching already installed. These cables carry an integrated DSP, retimer or gearbox to translate between the two lane PAM4 interface at one end and the four lane NRZ interface at the other, so the upgrade can be staged without stranding the existing fabric.

What comes next

The same pressure continues past 100GbE. 200G SFP112, 400G dual lane, QSFP112 and OSFP all push bandwidth up while holding power and thermal behaviour inside what a rack unit can dissipate. As AI, cloud and 5G buildouts continue, port density per rack unit will keep being weighed against interoperability and efficiency rather than treated as a secondary concern.

These four form factors are not competing for the same socket. QSFP28 holds the installed base. SFP-DD offers density without abandoning SFP+ and SFP28. DSFP serves compact telecom deployments. SFP112 points at where single lane 100G is heading. Together they give network designers room to build denser and more power efficient infrastructure without committing to one interface for every role.