Two and a half times the bandwidth in the same cage, at a similar cost per port - which makes the cost per gigabit far better.

QSFP+ delivered 40 Gb from four 10 Gb lanes. QSFP28 delivers 100 Gb from four 25 Gb lanes in the same mechanical format, using the same cages, the same cabling types and the same faceplate space.

For a switch designer that meant offering 100 Gb ports where 40 Gb ports had been, with no loss of density. For a network designer it meant an immediate capacity increase without changing the physical layout of the row.

The breakout relationship reinforced it. 100 Gb divides into four 25 Gb connections, and 25 Gb became the server access speed - so uplink and access speeds share a lane rate and the same cables serve both roles.

40 Gb, needing four lanes for less bandwidth and more fibre, had no answer to that, and its deployment fell away quickly once 25/100 equipment became widely available.

Parallel short-reach for in-row multimode, wavelength-multiplexed for existing duplex fibre, single-mode variants for campus and longer links.

Parallel short-reach modules use eight fibres of multimode through an MPO connector and reach around 100 metres, which covers within a row and between adjacent rows. They are the cheapest option where MPO trunking is installed.

Wavelength-multiplexed modules put the four lanes onto a single duplex pair, which lets 100 Gb run over existing duplex fibre with no cabling change - usually the deciding factor in buildings rather than purpose-built data halls.

Single-mode variants cover 500 metres, 2 kilometres, 10 kilometres and beyond depending on type, for links between buildings and between sites.

Parallel single-mode variants also exist and support breakout to four separate 25 Gb links over distance, which is useful where four remote destinations are fed from one port.

Match to the fibre you have before choosing on price.

The port is placed in breakout mode and appears as four independent interfaces, connected by a splitter cable to four SFP28 ports.

Most data centre switches support this on at least some ports. The configuration usually requires the port - and often its whole port group - to be reset, so it should be planned rather than done casually.

The cable is a direct attach assembly with one QSFP28 end and four SFP28 ends for in-rack connections, an active optical equivalent for longer runs, or an MPO-to-duplex fan-out with separate optical modules where distance requires it.

This is the standard way a spine or aggregation switch feeds servers or small switches: one port serves four devices at lower cost than four ports.

The planning points are that support may be limited to particular ports, that all four sub-links share one physical cable and one module, and that error correction settings apply per sub-interface and must match each connected device.

Generally yes, running at 40 Gb - which is what allows a 100 Gb switch to connect to an existing 40 Gb estate.

The format and cage are shared, and most platforms detect the module and set the port speed accordingly. That makes a phased upgrade straightforward: install 100 Gb capable switches, keep using 40 Gb optics against older equipment, and move links to 100 Gb as the far end is replaced.

Some platforms restrict which ports support the lower speed, and breakout behaviour can differ between 40 Gb and 100 Gb modes on the same port, so check the documentation for the specific switch.

The reverse is not possible - a QSFP28 module in a QSFP+ port cannot run at 100 Gb because the port's lanes are 10 Gb.

Where both speeds will be in use, plan the port allocation deliberately rather than mixing arbitrarily, since port groups sometimes impose constraints that only appear at configuration time.

Usually the stronger mode, and it must match at both ends - mismatches account for more failed 100 Gb links than anything optical.

At 25 Gb per lane the raw error rate is high enough that correction is designed into the link. The mode required depends on the module type and reach: short direct attach connections may need little, optical links generally need the stronger scheme.

Devices differ in what they enable by default, and where two ends disagree the result is either no link or a link with high error counts.

The diagnosis is unintuitive because the optics look perfect - transmit and receive powers are within specification and nothing indicates a problem.

So when commissioning 100 Gb, check the correction setting on both ends explicitly, particularly between different manufacturers' equipment. It is a configuration item that deserves to be in the build standard rather than discovered each time.

Considerably more than an SFP28, and in a densely populated switch the total is a real design constraint.

A 100 Gb module dissipates several watts, and longer-reach and coherent variants more. A switch with 32 such ports fully populated is dissipating a substantial amount from the module cages alone, in addition to the switch itself.

The practical consequences: check whether the platform supports full population with the intended module type, since some specify limits for higher-power modules; confirm the airflow direction suits the cabinet; and ensure the cabinet's cooling and the rack power budget account for it.

Modules report their own temperature, and reading those figures across a populated switch identifies thermal problems before they cause failures - modules at the hot end will show it first.

Direct attach copper dissipates far less, which is one reason in-rack connections use it in preference to optical modules wherever the reach allows.

QSFP-DD and OSFP for 400 Gb, both designed so that existing QSFP modules can still be used - which protects the investment in 100 Gb optics.

QSFP-DD doubles the electrical interface to eight lanes in a cage that remains backwards compatible: a QSFP28 module fits a QSFP-DD port and runs at 100 Gb. That means a 400 Gb switch can be deployed and populated with existing 100 Gb optics, upgrading links individually.

OSFP is a slightly larger alternative format with better thermal capacity, requiring an adaptor for backwards compatibility.

The practical guidance for anyone building at 100 Gb now is to check that the switch platform's port type offers a path forward, and to prefer structured cabling that supports higher lane rates - single-mode in particular, since it does not need replacing as speeds increase.

Multimode is where upgrades become expensive, because each speed generation reduces the supported distance on the same fibre.