SFP vs SFP+ vs SFP28 vs QSFP28: Types and Compatibility

Choosing the wrong pluggable transceiver - or assuming two form factors are interchangeable - is a classic source of link-down mystery. SFP, SFP+, SFP28, QSFP+ and QSFP28 share the same MSA (multi-source agreement) ecosystem but differ in speed, lane count and, critically, compatibility rules. This guide compares the five form factors and answers the questions that actually come up on site: can an SFP28 plug into an SFP+ port, and what happens if it does?

Form Factor Reference Table

SFP was introduced in 2001 to replace GBIC at up to 1G; SFP+ (SFF-8431) brought 10G to the same physical shell; SFP28 kept the shell for 25G single-lane; QSFP families quadruple the lanes for 40G/100G in one slot, which is why 4x SFP+ optics can often be replaced by a single QSFP+ module.

Lane Architecture: Why the Speeds Jump the Way They Do

The naming is arithmetic once you know the lane count. A cage is a mechanical and electrical container for a number of serial lanes, and the module's total capacity is lanes multiplied by per-lane rate:

• SFP / SFP+ / SFP28: one lane, one fibre pair. The shell stayed identical across three generations while the per-lane signalling rate moved from 1.25 Gbit/s to 10.3125 Gbit/s and then to 25.78125 Gbit/s. Line rates carry overhead, which is why a "10G" port runs at 10.3125 Gbit/s on the wire.
• QSFP+ / QSFP28: four lanes in one cage. QSFP+ uses 4 x 10G for 40G total; QSFP28 uses 4 x 25G for 100G. From the switch's point of view this is the key difference from the SFP family - a QSFP port can often be split into four independent slower ports.
• Higher tiers: QSFP56 (4 x 50G = 200G) and QSFP-DD (double density, eight lanes) extend the same idea, while OSFP uses a larger shell to handle the power and cooling budget of 400G/800G optics.

Encoding changed along the way too. Everything up to 25G per lane uses NRZ signalling - one bit per symbol. 50G per lane and above uses PAM4, four levels per symbol, which is why a 400G module is not simply "more lanes" and needs different electrical specifications (SFF-8024 identifiers and the CMIS management interface) than a 100G module.

Compatibility Rules That Matter

• 1G SFP in an SFP+ port: generally works - SFP+ ports are designed to accept SFP optics, and the link runs at 1G. Many 10G switches auto-negotiate down or need the port forced to 1G.
• SFP+ in an SFP port: do NOT do this - the reverse direction is not supported and can damage the module or port.
• SFP28 in an SFP+ port: the connectors and pinouts are mating-compatible, so a 25G SFP28 module inserted into an SFP+ (10G) port negotiates down and runs at 10G if the switch allows the speed. Conversely, a 10G SFP+ module in an SFP28 port works when the port is configured for 10G.
• QSFP28 vs QSFP+: same shell and management interface (SFF-8636); a QSFP28 port usually accepts QSFP+ modules at 40G, and QSFP28 100G modules do not drop into QSFP+ 40G ports.
• Breakout: QSFP28 ports commonly support breakout to 4x25G (or 4x10G) via DAC/fiber fan-out cables when the platform supports speed 4x25G port modes.

Read the rules as a hierarchy: mechanical fit is the weakest guarantee, electrical signalling is the second, and firmware permission is the third and most decisive. A module can mate perfectly, signal correctly, and still produce no link because the switch's port profile does not list that speed or that media type. Before blaming the optic, check whether the port is even capable of the rate you are asking for - a 100G module in a port whose controller tops out at 40G will never come up, no matter how many times it is reseated.

Speed Negotiation and Port Configuration

Most "the module is broken" cases are port-profile problems. Two examples show the pattern:

! Cisco-style: force an SFP+ port down to 1G for an SFP optic
interface TenGigabitEthernet1/0/1
 speed 1000
 no shutdown

! Huawei-style: allow 10G on an SFP28-capable port
interface 25GE1/0/1
 speed 10000
 undo shutdown

A few practical rules follow from how negotiation works:

• Both ends must agree. Forcing one side to a fixed speed while the other auto-negotiates is a common cause of a link that comes up and then flaps.
• FEC is part of the negotiation. 25G and 100G links generally require forward error correction, and some platforms want it explicitly enabled on one or both ends (for example fec rs versus fec fc on 100G). A mismatch presents as a link that never achieves sync, not as a completely dead port.
• Breakout mode is a port-level setting. Splitting a QSFP28 into four 25G ports renumbers the interfaces on some platforms and requires the module and cable to support the split; a plain 40G DAC cannot be broken out.

The Standards Behind the Modules

This matters operationally: the reason a switch sometimes reports an "unknown" or "unsupported" transceiver is that the module's EEPROM does not present the identifier and vendor strings the platform expects, not that the optics are faulty. Standards also explain compatibility in the other direction - because SFP+ management follows SFF-8472, a switch can read temperature, supply voltage, TX bias and RX power from a 1G module in a 10G cage just as it does from a native 10G optic.

Vendor Coding, EEPROM and Lock-in

Every pluggable carries a small EEPROM that answers three questions for the host: what type of module is this, who made it, and is the checksum valid. Vendors have long used the vendor field as a gate. When a switch rejects a third-party module it is usually not seeing a physical problem at all; it is declining to bring up a port whose EEPROM is not in its approved list.

• Compatible/coded third-party optics carry a programmed EEPROM that reproduces the expected vendor string and checksum. They work on most platforms but can be rejected after a firmware upgrade that tightens the allow-list.
• Uncoded generics may work on less restrictive vendors, and are the usual cause of "works in the lab, refused on site".
• DAC and AOC cables are programmed the same way, so a cable that works between two identical switches may be refused on a different model.

Reading the EEPROM is the fastest diagnostic: show interface transceiver (Cisco), display interface transceiver verbose (Huawei), /interface ethernet monitor sfp1 (MikroTik) or ethtool -m (Linux) will either show details or print an error that tells you the module was not recognised at all.

DAC, AOC and Optics

Not every pluggable is an optical module. DAC (direct attach copper) cables have transceivers molded into the ends - lowest cost and power for short racks (up to ~7m at 10G); AOC (active optical cable) extends the same plug-and-play model to tens of meters; standalone optical modules need fiber and are chosen by distance and fiber type (SR over multimode OM3/OM4, LR over single-mode OS2).

The DAC family splits further, and the split matters when a cable does not work:

• Passive DAC - no electronics in the cable, cheapest and lowest latency, but length-limited (roughly 3-5 m at 25G, up to 7 m at 10G).
• Active DAC (ACC) - signal conditioning in the cable ends, extending passive copper to around 5-10 m.
• Active optical cable (AOC) - a permanent fibre pair with fixed optics; 10-100 m typical, longer than any copper option and lighter than separate optics plus patch cords.
• Breakout DAC/AOC - one QSFP end, four SFP/SFP+ ends, for connecting a 100G switch port to four 25G servers.

Fiber Types and Reach Classes

The reach suffix in a module part number tells you the fibre it expects. Getting it wrong is the second most common cause of a link that comes up with errors or not at all.

Two rules: never mix SR and LR at the two ends of a link (the wavelengths do not match), and never assume a longer-reach module will "just work" over a shorter run - a 40 km ER module over 100 m usually needs an inline attenuator to avoid overloading the receiver.

Breakout, Channelization and Port Modes

Breakout is the feature that gets the most use in modern top-of-rack designs. A 100G QSFP28 port can be presented as four 25G ports, or a 40G QSFP+ port as four 10G ports, when the switch supports the mode and the module or cable supports the split. Practical points:

• The breakout cable must physically split the lanes to four LC or MPO connectors; a standard duplex module cannot be broken out.
• Some platforms require the breakout to be configured before the module is inserted, and interface numbering changes afterwards.
• Channelized QSFP28 ports that are split cannot be combined with the module's own 100G operation at the same time.
• Both ends must agree: breaking out one side while the other side expects a single 100G link gives four dead ports.

Digital Optical Monitoring and Link Budget

Most optics above 1G expose digital optical monitoring (DOM/DDM): TX power, RX power, temperature, supply voltage and laser bias current. These five numbers turn a vague "the link is flapping" into a specific finding.

• RX power far below the module's sensitivity - dirty connector, damaged patch cord, wrong fibre, or excessive loss from too many splices.
• RX power above the saturation point - the link is too short for the reach class; add an attenuator.
• TX power normal, RX power absent - the far end is not transmitting, the fibre is broken, or the pair is crossed.
• Temperature climbing steadily - airflow or a failing module; high-temperature optics fail intermittently before they fail completely.

Read the values against the module's datasheet rather than against "looks reasonable", and compare a suspect port with a known-good port of the same type on the same switch.

Newer Form Factors You Will Meet

The five classic form factors are still the bulk of installed base, but new deployments increasingly include:

• QSFP56 / SFP56 - 200G/50G per-lane PAM4 versions of the familiar shells.
• QSFP-DD - eight lanes in a shell that also accepts a QSFP28 module in the same cage, which makes it the common 400G stepping stone.
• OSFP - a slightly larger shell for 400G/800G with more power and thermal headroom; it does not fit a QSFP-DD cage.
• SFP-DD - two lanes up to 25G each for 50G in an SFP-sized shell.

Practical Selection Advice

• Match the module's reach class to the link: don't pay LR prices for a 5 m rack link.
• Keep the same vendor/standard on both ends (SR-SR, LR-LR) and check DOM (digital optical monitoring) support if you plan link-budget monitoring.
• When a link fails after a swap, first suspect a form-factor or speed mismatch, then verify with the steps in our SFP transceiver troubleshooting checklist.

A selection checklist that survives contact with procurement and with the site:

1. Port speed and cage type on both ends (SFP/SFP+/SFP28/QSFP+/QSFP28 and the port profile the switch actually supports).
2. Link distance and fibre type available between the two racks.
3. Media: optical module plus patch cords, or DAC/AOC for in-rack links.
4. Coding: does the platform reject uncoded modules, and is the cable/optic coded for both vendors?
5. Reach class and, for short links with long-reach modules, an attenuator.
6. DOM support if link-budget monitoring is required.
7. Spares of the same part number - a stock of "equivalent" optics from three vendors makes fault isolation harder, not easier.

Troubleshooting: When a Link Does Not Come Up

Common Myths

• "All SFP cages are the same." The cage is the same; the host's supported rates and firmware allow-list are not.
• "A 25G module in a 10G port is dangerous." It is usually harmless and negotiates down - but check the platform, because some designs do not support the fallback.
• "Longer reach is always better." Over-specified optics need attenuation and cost more; they do not add margin.
• "A copper DAC is always cheapest." Only for very short runs - beyond a few metres the higher-speed copper options approach optical pricing while staying length-limited.

Choosing the optic is only half the job - the fiber plant behind it decides reach: see our CWDM vs DWDM comparison when wavelengths are involved, and the MTU mismatch troubleshooting guide for the frame-size issues that show up right after new optics go in.

原文链接：https://www.fs.com/eu-en/blog/sfp-sfp-sfp28-qsfp-qsfp28-what-are-the-differences-406.html