A few metres, falling as the lane rate rises - roughly five to seven metres at 10 Gb per lane, around three at 25 Gb, and less at 50 Gb.

The signal travels electrically over copper, and higher frequencies attenuate faster. Each generation of lane rate therefore reduces the usable length.

Passive cables contain nothing but the conductors and are limited by that attenuation. Active cables include signal conditioning electronics that extend the reach - roughly doubling it in many cases - while consuming some power and adding a small cost.

What follows in practice is that direct attach is an in-rack technology. Server to top-of-rack switch is the standard use; rack to adjacent rack is sometimes possible; anything further needs active optical cables or transceivers.

Measure the actual routed length rather than the straight-line distance. Cable management, vertical runs and service loops add significantly, and a cable that is theoretically long enough may not reach once it is routed properly.

Cost, power, latency and reliability - it wins on all four within its reach.

Cost is the obvious one: a direct attach cable costs a fraction of two transceivers plus a patch lead, and across a data centre with thousands of server connections that is a large sum.

Power matters at scale. An optical module pair consumes several watts per link; a passive copper cable consumes essentially nothing. Across a full rack the difference is measurable in the cooling load as well as the electricity.

Latency is lower because there is no optical conversion at either end - a small difference, but meaningful for latency-sensitive workloads.

Reliability benefits from there being no end faces. Contamination is far and away the commonest cause of optical link problems, and a sealed copper assembly cannot suffer from it.

The trade is reach and rigidity. Within a rack, direct attach should be the default; beyond it, optics are the answer.

Often, but the module ends carry coding and both switches must accept theirs - which makes a cable between two vendors' equipment a specific thing to order.

Each end presents itself to its switch as a module, complete with vendor identification. A switch that enforces coding will reject an end it does not recognise.

For a cable between two different manufacturers' switches, the two ends must be coded differently. Reputable optics suppliers sell exactly this and will code each end appropriately - state both platforms when ordering.

Beyond coding, electrical interoperability is generally good since the interfaces are standardised, but compatibility lists are worth checking for high lane rates where signal integrity margins are tighter.

The practical advice is to order from a supplier who will confirm the combination, and to test one before buying a hundred. A cable that will not link is inexpensive to replace individually and expensive to replace in bulk.

They are thick, stiff and have a large bend radius - and a rack of them is genuinely difficult to route neatly.

A twinaxial assembly is much less flexible than a fibre patch lead, and it does not tolerate tight bends. Forty of them leaving a top-of-rack switch is a substantial physical bundle that resists being tidied.

The consequences are practical. Airflow can be obstructed if the bundle sits across an exhaust. Access to individual ports becomes difficult once the bundle is dressed. And a cable bent past its radius during installation can suffer intermittent faults that are hard to attribute.

Mitigations: choose lengths accurately rather than using long cables and coiling the excess, use horizontal cable management to break the bundle up, route away from airflow paths, and consider thinner higher-gauge cables where the reach allows - the gauge number rises as the cable gets thinner and the reach falls.

Active optical cables are much thinner and lighter, which is sometimes reason enough to prefer them even within reach of copper.

Passive for short in-rack connections; active where the reach is beyond passive but still short enough to avoid optics.

A passive cable is conductors and connectors, with no electronics. It consumes no power, adds no latency, cannot fail electrically, and is the cheapest option.

An active cable includes signal conditioning at the ends to compensate for attenuation, roughly doubling the reach. It draws a small amount of power and costs more, but remains far cheaper than a pair of optical modules.

The choice is usually decided by measurement: if the routed length is within passive reach, use passive.

One further consideration is host tolerance. Some switch platforms have limited support for longer passive cables at higher lane rates, and specify active for anything beyond a short distance. Check the platform's supported cable list, because this is a case where a cable that physically fits may not be supported.

Yes, and breakout assemblies are among the commonest uses - a quad end at one side splitting into four single-lane ends at the other.

A QSFP28 to four SFP28 assembly connects one 100 Gb switch port, configured in breakout mode, to four 25 Gb server connections. The same pattern exists at other speeds.

Within a rack this is the standard way to feed servers from an aggregation port, and it saves both port count and cost.

The considerations are the same as any breakout: the switch port must support the mode, the configuration typically requires a port group reset, and all four links share one physical cable - so a single damaged cable removes four connections.

Lengths are fixed and the split point is at the far end, so the assembly must be chosen to suit the physical layout. Measure before ordering, because a breakout cable that is too short cannot be extended and one that is too long leaves four tails to manage.

By substitution, mainly - there are no optical diagnostics to read, so the switch's error counters and a known-good cable are the tools.

The module ends do report basic identification and some report temperature, but there is no equivalent of optical receive power to measure.

So the diagnostic sequence is: check the port's error counters at both ends, confirm the speed and error correction settings match, reseat both ends, then substitute a known-good cable.

Physical inspection is worth doing before substitution. A cable bent past its radius, crushed by a cabinet door, or strained at the connector is a common cause, and the damage is often visible.

The sealed construction means most failures are mechanical rather than electrical, and they occur at the point where the cable meets the module housing - which is why cables should be supported so their weight does not hang on the connector.

Keep spares of each length and coding combination in use, since a specific cable is not something that can be improvised.