25 Gigabit Ethernet is the workhorse of the modern leaf layer. It is the rate most new server NICs ship at, it is the building block of 100G (4 × 25G), and it is what 5G fronthaul runs on. If you are designing a switch, a NIC, or a network appliance with SFP28 ports, the cage is the component that decides whether those ports perform.

The good news: an SFP28 cage is mechanically the same animal as an SFP+ cage. The MSAs were designed for that. The bad news: mechanically compatible does not mean electrically equivalent, and 25G is the rate where the difference starts to show.

This guide covers what to check — mechanical reuse, signal integrity, thermal, EMI, and port density — when you specify an SFP28 cage. For the broader "which cage for which rate" question, see SFP vs SFP+ vs SFP28; this article goes deeper on the 25G case.

1. What SFP28 actually is

SFP28 is the 25G member of the SFP family: one electrical lane at 25.78125 GBd, NRZ signalling. The relevant IEEE standard is 802.3by, which defines 25GBASE-CR and 25GBASE-CR-S (twinax copper, DAC), 25GBASE-KR (backplane), 25GBASE-SR (multimode fibre), and the 25GBASE-LR/ER variants came through the same family of work.

The important structural facts:

• 25G is one lane, not four. 100G is 4 × 25G, which is why a QSFP28 port breaks out to four SFP28 ports so cleanly.
• SFP28 modules are mechanically interchangeable with SFP and SFP+. Same shell, same latch, same 20-position host connector interface. A 25G module will physically insert into a 10G cage.
• The host connector and cage must be specified for the higher rate. The SFP+ MSA (SFF-8431 electrical) was written for 10G; SFP28 requires the host interface to behave to roughly 12.9 GHz — the Nyquist frequency of a 25.78125 GBd NRZ signal — and typically a bit beyond, to 14 GHz or so, to cover the harmonics that matter for eye closure.

That third point is the whole article, really. The mechanical interchangeability is a feature for operations teams and a trap for hardware teams.

2. Mechanical: yes, you can reuse the footprint — with conditions

If you have an existing 10G SFP+ design, the SFP28 cage will drop into the same PCB land pattern and the same faceplate cutout. That is by design. What you must confirm before reusing:

The condition: reusing the footprint is fine; reusing the part number is a decision that needs data. A cage that was characterised to 10 GHz may or may not be characterised to 14 GHz. Ask for the S-parameter model and the compliance report at 25.78125 GBd.

3. Signal integrity: the parameters that matter at 25G

At 10G NRZ, the cage-plus-host-connector assembly was a modest part of the channel budget. At 25G NRZ, it is a meaningful discontinuity. Design targets you should be holding the supplier to (typical industry practice — always check against your SERDES vendor's budget):

What to actually ask for:

1. Touchstone S-parameters (.s4p or .sNp) of cage + host connector + paddle card, measured, not simulated-only.
2. Model-to-measurement correlation data. A model nobody has checked against a VNA is a marketing artifact.
3. A compliance report at 25.78125 GBd — not "tested to 10G, extrapolated to 25G."
4. The recommended paddle-card stack-up and trace geometry. This is where most 25G designs actually go wrong. The paddle card is a controlled-impedance structure and it is usually the vendor's reference design; deviating from it without re-simulation is how you end up with a port that links but has no margin.

Paddle card note: most SFP28 cages are supplied with, or designed against, a specific paddle-card design. The paddle card carries the 25G differential pairs from the host connector to the host PCB. Its stack-up, trace width, spacing, reference planes and length are all part of the channel. Treat it as part of the cage, not as a separate mechanical detail.

4. Thermal: 1.5 W is the number, but 48 ports is the problem

SFP28 module power, by type:

Against SFP+ SR at ~1 W, that is a 50% increase in per-port heat at the same faceplate density. In a 48-port 1U chassis it is the difference between 48 W and 72 W coming off the cage bank — before you count the ASIC.

Cage-level thermal decisions:

• Gap pad / thermal interface material between the module top and the cage inner wall. This is the cheapest and most effective intervention. Watch the compression set; a pad that bottoms out over the product life stops working.
• Heatsink options. Standard SR modules at 1.5 W usually do not need a heatsink in a well-designed 1U. LR and ER modules, and any design at high inlet temperature, often do. Cage designs that accept a "riding heatsink" — a finned block that sits on the module and is retained by the cage — are the usual answer.
• Airflow. Confirm the direction (port-side intake vs exhaust) and the inlet temperature class (ASHRAE A1–A4). At A3/A4 inlet temperatures, a 1.5 W port is a different design problem.
• Measure with DOM. SFF-8472 digital diagnostics give you the module's internal temperature over I2C. Log it during thermal validation; it is free and it is ground truth.

For the full treatment, see SFP Cage Thermal Management.

5. EMI: higher frequency means a leakier cage

A 25G NRZ signal has energy at 12.9 GHz and above. RF leakage through the cage-to-bezel and module-to-cage interfaces is more efficient at 12.9 GHz than it is at 5 GHz. The practical consequences:

• EMI spring fingers or an integrated gasket at the bezel interface are effectively mandatory for 25G, not optional.
• Dust/EMI plugs in unpopulated ports. An empty SFP28 port is a slot antenna. This is the single most common cause of a first-pass emissions failure on a new switch.
• Cage-to-chassis bonding. The cage must be at chassis ground at RF, not just at DC. Multiple grounding tabs, short paths, and good contact with the bezel or chassis rail.
• Light pipe EMI. A plastic light pipe through the front of the cage is an aperture. Use a metallised or shielded light pipe assembly, or route the light through a waveguide-below-cutoff aperture.

For a full treatment of the options, see SFP Cage EMI Shielding.

6. Where SFP28 gets used — and what that implies for the cage

Top-of-rack (ToR) / leaf switch. 25G to the servers, 100G to the spine. 48 × SFP28 in 1U plus 6–8 × QSFP28 uplinks is the archetypal configuration. Implication: belly-to-belly or 1×6 ganged cages, EMI-tight, with thermal headroom for 1.5 W on every port.

100G breakout. One QSFP28 port on the switch, four SFP28 ports on the servers, via a breakout DAC or AOC. Implication: mechanical and latch robustness matters more than anything, because breakout cables get pulled and re-seated far more often than structured cabling.

Server NICs. 25G SFP28 on the NIC, often 2 ports. Implication: low-profile cages, and often a cage designed to fit under a PCIe bracket with tight airflow.

5G fronthaul (eCPRI). 25G SFP28 between radio units and baseband, frequently industrial temperature range and often outdoor-adjacent. Implication: industrial-temperature cages (−40 to +85 °C), corrosion-resistant plating, and EMI plugs because the far end is a tower.

Storage and appliances. Mixed 10G/25G, often 1×1 or 1×2 cages. Implication: cost sensitivity, and the temptation to reuse a 10G cage — resist it unless you have the 25G data.

7. Form factors: 1×1 through belly-to-belly

Ganged cages improve EMI (no gap between ports at the bezel) and speed assembly, but they concentrate heat: the middle ports of a ganged assembly have neighbours on both sides. Model it.

See SFP Cage Port Configuration for a detailed look at the trade-offs.

8. Press-fit vs solder tail for SFP28

The same arguments that apply at 10G apply here, with one addition:

• Press-fit is preferred for high-reliability and for boards that see multiple reflow cycles (paddle-card rework, second-side assembly). It also gives a stronger mechanical connection for the heavier 25G cable assemblies. See Press-Fit SFP Cage.
• Solder tail is lower cost and fine for 1×1 designs and for products that will never be reworked.
• The addition at 25G: signal integrity. Press-fit tails add a small but real discontinuity at the via. If you are tight on IL budget, ask the vendor for the press-fit and solder-tail S-parameters and compare.

SFP28 cage selection checklist

FAQ