For 25G access, confirm FEC mode and breakout configuration before optics: most links that fail at commissioning fail for configuration reasons, not optical ones.
01The connection brief
A 25G server access link joins a network adapter in a server to a leaf or top-of-rack switch port. This page explains how that link is planned and specified. It is an application concept written as a reference, not a description of a delivered build, and it contains no customer, quantity shipped or measured result.
The brief starts with the adapter side, because server network adapters are less forgiving than switches about what they will accept and how they negotiate. Record the adapter model and firmware, the switch model and software release, the rack layout, and whether the link stays inside one rack or crosses to an end-of-row position. Those four facts decide almost everything that follows.
02The design inputs
Document the adapter and its supported media list, the switch port type, and whether the switch port is a native SFP28 cage or a lane broken out from a 100G QSFP28 port. Breakout changes both the cabling and the configuration, so it belongs in the inputs rather than being discovered later.
Record the physical distance from server to switch as cable route length, including the vertical path through cable management. Record the fiber grade if fiber is used, and the cable category and gauge if copper is used. Note the forward error correction expectation on both sides, since 25G has several FEC modes and a mismatch prevents the link from coming up. Finally capture rack power and airflow, because access layers are dense.
03The topology and optical path
Three arrangements dominate. In-rack server to top-of-rack switch is usually direct attach copper or a short fiber jumper. Server to end-of-row switch is a structured fiber run through a patch panel. Breakout from a 100G switch port to four 25G server ports uses an MPO to LC breakout assembly or a breakout direct attach cable, and the switch port has to be configured for breakout mode before any of it works.
Write the path element by element in every case. Even a short in-row fiber link passes through at least two mated connector pairs at the panel, each counted at 0.3 to 0.75 dB, and those pairs are what turn a nominally comfortable multimode budget into a marginal one.
04Product families to evaluate
For in-hall multimode links the 25GBASE-SR SFP28 850nm 100m VCSEL Multimode Duplex LC module is the standard family, rated to 100 m on OM4. Where the run leaves the hall or uses installed single-mode, the 25GBASE-LR SFP28 1310nm 10km Single-Mode Duplex LC module is the usual choice, with the 25G SFP28 ER 1310nm 40km Single-Mode Duplex LC Transceiver available for longer routes.
Single-fiber routes are covered by the 25GBASE-BR20 SFP28 BiDi Single-Fiber 1270/1330nm 20km WDM module. Where some servers remain at 10 gigabit during a transition, the 10G SFP+ families handle those ports in the same switch, and the 10G SFP+ Copper RJ45 10GBASE-T 100m Cat6a/Cat7 modules cover any hosts that are still copper attached.
05Link budget and reach considerations
25GBASE-SR reach is set by modal bandwidth and by the FEC mode in use, and the practical planning figure for OM4 is 100 m, with shorter distances on OM3. The channel insertion loss allowance in the module datasheet is the number to compare against, so count mated pairs at 0.3 to 0.75 dB and splices at about 0.1 dB and confirm the total fits.
Single-mode links are budgeted arithmetically using roughly 0.35 dB/km at 1310 nm, plus connector and splice losses, against the difference between minimum launch power and receiver sensitivity. Keep about 3 dB of unallocated margin. On very short single-mode links, check the maximum receiver input power as well, since an ER-class module across a rack row can overload the far receiver.
06Environment and power
Access layers are the densest part of a data hall, so thermal planning is the practical constraint. SFP28 optics in these families are commercial grade, rated 0 to 70 degrees Celsius at the case, and a fully populated switch in a hot aisle is the realistic worst case rather than the room temperature reading.
Power differences between media matter at scale. A passive direct attach cable draws essentially nothing, a short-reach optic draws more, and a long-reach single-mode optic more again. Across a full rack of servers those differences add up in both the switch power budget and the heat load. Where server and switch sit in the same rack, direct attach is usually the lowest power and lowest cost option available.
07Validation and acceptance
Acceptance for 25G should focus on the error counters rather than on link state, because a marginal 25G link will come up and then accumulate corrected errors quietly. Confirm that both ends agree on the FEC mode, then read the pre-FEC bit error rate or corrected codeword count if the platform exposes it, and record the value.
On fiber links, read digital diagnostics at both ends and compare received power with the calculated budget. On copper direct attach, confirm the cable length and gauge match the configuration expectation. Run traffic for a defined soak period, confirm that uncorrected errors stay at zero, and store the counters with the as-built record so a later drift can be recognised as drift.
08Procurement and BOM
Write the BOM per server port and multiply by the port count, then add the switch-side quantity separately, because a breakout arrangement makes the two sides different. Four 25G server links fed from one 100G port need four server-side modules or cable ends but only one switch-side transceiver or breakout cable.
List the adapter-side and switch-side coding requirements separately, since the accepted media list of a server adapter is not the same as that of a switch and coding is confirmed per platform. Include jumpers, breakout assemblies and any attenuators identified by the budget calculation, each with its length and connector style. Spares are a policy figure per optic type rather than a fixed count, sized against how quickly a replacement can reach the hall.
09Common mistakes
The most common failure is a FEC mismatch between adapter and switch, which produces a link that never comes up despite healthy optics on both sides. The second is forgetting to configure breakout mode on the 100G switch port before connecting a breakout assembly.
The third is planning multimode reach from the headline figure without counting patch panel connector pairs. The fourth is mixing OM3 and OM4 in one channel and budgeting as if the whole run were OM4. The fifth is ordering direct attach cables by nominal rack distance rather than by routed length, which produces cables that are physically too short once cable management is respected. The sixth is assuming a server adapter accepts whatever the switch accepts.
10Next step
Assemble the adapter inventory with models and firmware, the switch inventory with models and software releases, the rack layout with routed distances, and a statement of which ports will be native and which will come from breakout.
Then add candidate items from the 25G SFP28 and 10G SFP+ listings to an inquiry list, separating server-side and switch-side quantities so that the two coding requirements stay visible as distinct lines. The project worksheet provides blank fields for the same inputs. Because this page is a planning concept rather than a delivered design, the media choice, the breakout plan and the FEC expectation should all be confirmed against the actual adapter and switch documentation before anything is ordered or installed.
11Design inputs to document
| Design input | What to record |
|---|---|
| Adapter inventory | Server network adapter model, firmware and supported media list. |
| Switch inventory | Model, software release, and native versus breakout port plan. |
| Routed cable length | Real path length including vertical cable management, per link. |
| Media grade | OM3, OM4 or single-mode fiber, or copper gauge for direct attach. |
| FEC expectation | The forward error correction mode required at each end. |
| Mated pair count | Patch panel connections present in each fiber channel. |
| Rack thermal and power | Airflow direction and remaining switch port power budget. |
| Acceptance criteria | Soak duration and the error counter thresholds that define a pass. |
12Planning bill of materials
| Item | Configuration | Quantity |
|---|---|---|
| 25G SFP28 SR transceiver | 850 nm VCSEL, duplex LC, OM3 or OM4, 100 m class | 2 per in-hall link |
| 25G SFP28 LR transceiver | 1310 nm, duplex LC, single-mode, 10 km class | 2 per single-mode link |
| Duplex LC jumper | Fiber grade and length to match the routed path | 2 per fiber link |
| 100G to 4x25G breakout assembly | MPO-12 to LC or breakout DAC, length to specify | 1 per four-port group |
| Spare optics | Same optical class and coding as production | Percentage to specify |
Planning quantities only. The final bill of materials is confirmed with your quotation.
13Frequently asked questions
Why does a healthy 25G link fail to come up?
Most often because the two ends expect different forward error correction modes. 25G defines more than one FEC option and some adapters and switches default differently. Check and pin the FEC setting on both sides before investigating the optics. A second common cause is a 100G port used for breakout that has not been configured into breakout mode.
What reach should be planned for 25GBASE-SR?
Take 100 m on OM4 as the working figure, with less on OM3, and then verify that the total connector loss in the channel fits the datasheet allowance. A link that passes through two patch panels can carry over a decibel of connector loss before any fiber attenuation, which is enough to matter at the top of the reach range.
Is direct attach copper suitable for server access?
Inside one rack it usually is, and it draws the least power of the available options. It is a fixed-length assembly, so it cannot be patched through a panel and must be ordered by routed length rather than by rack-to-rack distance. Beyond the rack, a fiber jumper or a structured fiber run through panels is the practical choice.
Do server adapters need the same module coding as switches?
Not necessarily. Adapters and switches read the module EEPROM independently, and each may expect different coding or maintain a different accepted media list. Coding is therefore confirmed per platform on both sides rather than assumed from one of them. The planning BOM lists server-side and switch-side modules as separate line items even when the optical specification on each is identical.
Technical review note: confirm model-specific specifications with the product datasheet and the intended host documentation. Product images and application photographs are illustrative.
Discuss this with our team
