Inside a single rack, a passive DAC is almost always the right answer: no active silicon, no FEC negotiation to nurse, and the lowest power and parts count per port at any speed up to 400G. Between roughly 5 and 100 metres, an active optical cable (AOC) is usually the cheapest assembly that will actually hold a link, which makes it the default for rack-to-rack and row-to-row runs. Past 100 metres, or wherever you have patch panels, single-mode fibre or a need to replace one end of a link without pulling the whole run, you buy pluggable transceivers and patch cords. These three are not really competing on price; they compete on distance, and distance decides the answer before price gets a vote.
Start With Distance, Then Work Backwards
The reason cabling decisions go wrong is almost never a technology mistake. It is a measurement mistake. People size a run from the front of the switch to the back of the server and forget the service loop, the vertical cable manager, the overhead tray, and the fact that a cable routed around a PDU takes a longer path than the straight line they paced out. Add 0.5 to 1 metre of slack at each end before you look at a datasheet, and measure the longest run in the rack rather than the average. In our experience a rack that “needs 2 metre cables” usually needs 3 metre cables, and a rack that “needs 5 metre cables” is really an AOC job.
The second input is how many links sit in each distance band. A 42U rack with 32 servers on dual-rail 100G might produce 64 links under 3 metres and two links at 25 metres. Those two long links should not drag the entire bill of materials into the optical world, and the 64 short links should not be forced onto optics for the sake of standardisation. Segment the count by band first, then assign media per band.
What Each Option Physically Is
Passive DAC is a fixed assembly of twinax copper pairs terminated into pluggable shells at both ends, with an EEPROM in each shell so the port can read its capabilities. There is no retimer, no laser and no power draw worth modelling. The trade is mechanical: to keep insertion loss inside the channel budget, longer lengths need heavier gauge, and 24 to 26 AWG twinax is stiff, thick, and has a bend radius that punishes tight racks. Short 30 AWG assemblies route far more easily but will not reach as far.
Active copper (ACC or LACC) keeps the copper pair but adds a linear amplifier or redriver in each shell. That buys reach without moving to optics: on the 400G OSFP line NVIDIA specifies roughly 3 m for passive copper and 3 to 5 m for the active copper variant, and on 400G QSFP-DD the active copper family runs from about 3 m out to 100 m. Power stays in the low single-digit watts per link, and latency stays effectively copper-grade.
AOC is two optical modules permanently terminated onto a multimode ribbon at the factory, sealed and sold at a fixed length. It behaves like a cable but costs like a pair of optics, minus the patch cord and minus the labour of mating two connectors. Jacket diameter is around 3 mm and the bend radius is far friendlier than twinax, which is why AOC is what people reach for when a rack is already full of cable.
Transceivers plus patch cords split the link into three replaceable pieces: a module in the switch, a module in the NIC, and a fibre between them. That modularity is the whole point. It is also the only option that supports single-mode reach, structured cabling through a patch panel, and swapping one failed half of a link during a maintenance window instead of re-routing a cable through a live rack.
Reach and Media: The Comparison Table
The numbers below follow NVIDIA’s published LinkX reach figures for the Ethernet families and the standard MSA reaches for the optical modules. Treat the copper ceilings as hard limits, not targets.
| Assembly | Practical reach | Media | Power per link | Field replaceable | Best fit |
|---|---|---|---|---|---|
| Passive DAC | 100G QSFP28 up to 5 m; 200G QSFP56 up to 2.5 m; 400G QSFP-DD 2.5 m; 400G OSFP 3 m | Twinax copper | Effectively zero | Whole cable only | In-rack, ToR to server, GPU tray to leaf |
| Active copper (ACC/LACC) | 400G OSFP 3 to 5 m; 400G QSFP-DD 3 to 100 m | Copper with redriver | Low, single-digit watts | Whole cable only | Dense racks that need a little more reach than passive allows |
| AOC | 100 m on multimode at 100G; 100 m for 400G QSFP-DD; 50 m for the 400G OSFP multimode family | Multimode ribbon | Roughly 1 to 4.5 W for the whole cable | Whole cable only | Rack to rack, row to row, 5 to 100 m |
| Multimode transceiver plus patch | 100G SR4 and 200G SR4 at 100 m on OM4; 400G SR8 at 100 m on OM4 or OM5 | OM3 / OM4 / OM5 | Roughly 2.5 to 3.5 W per module at 100G | Yes, per end | Patch panels, mixed vendors, structured cabling |
| Single-mode transceiver plus patch | 500 m DR, 2 km FR, 10 km LR, 40 km ER | OS2 single-mode | Higher, and materially higher at 400G | Yes, per end | Campus, inter-building, long haul inside a facility |
What Changes on the Wire: Latency, Power and FEC
A passive DAC’s only contribution to latency is flight time, roughly 5 nanoseconds per metre. A 3 metre DAC is therefore about 15 nanoseconds, which is noise next to anything the application does. An optical link adds serialisation, deserialisation and clock recovery at both ends, which lands in the tens to low hundreds of nanoseconds per hop depending on the module and whether FEC is engaged. For storage replication across a campus, that matters. Inside a rack, it does not.
Power is where the three options diverge sharply. Copper draws essentially nothing and generates no heat inside the connector cage. A 100G multimode module sits in the 2.5 to 3.5 W range; a 400G module is an order of magnitude above that. Multiply by port count and the optics become a cooling problem: 64 active ports at 10 W each is 640 W of heat inside one switch chassis, which changes the airflow design of the rack, not just the electricity bill. A rack built entirely on DAC has no such penalty, which is why the cheapest cabling is often also the coolest.
FEC is the third difference and the one that causes the most support tickets. 100GBASE-SR4 mandates Reed-Solomon FEC in the standard, and 200G and 400G PAM4 links effectively always run it. A 100G passive DAC frequently comes up without FEC at all, which is part of why copper links negotiate faster and tolerate older firmware better. When one end insists on FEC and the other does not, or the two ends disagree on the clause, the port will show light, report a partner, and never pass traffic. If a new optical link refuses to come up while the DAC it replaced worked fine, check FEC and auto-negotiation before you suspect the module.
Passive Copper: Where It Wins and Where It Stops Working
Passive DAC wins on the metrics that matter inside a rack: no power, no heat, no cleaning, no FEC tuning, and the lowest part count per port. It is the correct choice for ToR-to-server links, for GPU trays cabled to a leaf in the same cabinet, and for lab benches where links get re-patched daily.
The ceilings are real. 100G QSFP28 passive copper tops out around 5 m, 200G QSFP56 at about 2.5 m, 400G QSFP-DD at 2.5 m and 400G OSFP at 3 m. Those are qualified lengths, not marketing suggestions; pushing past them produces links that train, drop, retrain and then quietly degrade a training job’s throughput rather than failing cleanly. If you need 6 m at 100G, the answer is active copper or AOC, not a longer DAC.
The second limit is mechanical. A 32-port switch fully populated with 26 AWG assemblies has a rear that looks like a bundle of garden hoses, with bend radii that fight the cable manager and airflow that has to go somewhere. Dense 1U platforms such as a 32-port QSFP28 SN2700 CS2ROS are happiest when the short links use thin-gauge assemblies and the long ones leave the rack entirely on fibre.
Active Optical Cable: The Middle-Distance Default
AOC occupies the band where copper has run out and discrete optics are over-engineered. At 100G, a single AOC reaches 100 m over multimode, which covers essentially every rack-to-rack run inside a hall. At 400G the QSFP-DD AOC family also runs to 100 m, while the 400G OSFP multimode family is specified to 50 m, so the exact ceiling depends on the port type you are cabling, not on the word “AOC”.
The operational advantages are physical. A 3 mm jacket with a tight bend radius is dramatically easier to route through a full rack than twinax, and it weighs a fraction as much, which matters when 64 cables hang off a single line card. The endfaces are sealed at the factory, so there is no inspection and no cleaning step, and no chance of a technician leaving a contaminated connector in a port. In dusty or high-traffic environments that alone justifies the price difference.
The trade is that length is fixed at order time and failure means replacing the whole assembly. An AOC cannot be re-terminated, cannot be shortened, and cannot be re-used when a rack gets re-arranged. Treat AOC as a consumable with a spare ratio, not as building infrastructure, and it behaves predictably.
Transceivers and Patch Cords: When Modularity Pays Back
Move to pluggable optics when any of four conditions is true: the run exceeds 100 m, the plant is single-mode, the link passes through a patch panel or cross-connect, or you need to replace one end without touching the other. Add a fifth: when the same physical fibre has to serve ports of different speeds over its life, because swapping a module is a two-minute job and swapping an AOC is a re-cable.
The cost structure is different rather than simply higher. You buy two modules and one patch cord instead of one assembly, and you also buy the discipline that optics demand: inspection scopes, one-click cleaners, and a rule that nothing gets mated without being inspected. Contaminated endfaces are the single most common cause of intermittent high-speed links, and they are invisible until you look. Teams that standardise on optics usually end up with a cleaning kit on every row and a spare module per speed per rack.
This is also the only route that scales cleanly across a mixed fleet. A 100G QSFP28 adapter such as the ConnectX-5 MCX516A and a 100G/200G QSFP56 adapter such as the ConnectX-6 MCX653106A can share the same physical fibre plant if the modules at each end match the port speed, which is exactly the flexibility a fixed-length cable removes.
Fiber Plant Prerequisites Before You Buy a Single Optic
Optics are only as good as the glass behind them. Before ordering modules, confirm the fibre grade, the connector type and the polarity scheme, because all three are decided by the plant you already own.
Grade first. 100G SR4 reaches 100 m on OM4 but only about 70 m on OM3, and 400G SR8 wants OM4 or OM5 at 100 m with low-loss connectors. If the installed plant is OM3 and the longest run is 85 m, the honest answer is a shorter reach or a different topology, not a module that claims 100 m. Second, connector and lane count: 100G SR4 uses an MPO-12 with eight fibres active, 400G SR8 uses MPO-16, and the single-mode DR and FR families use either an MPO-12 or an LC duplex pair. Third, polarity: MPO trunks come in type A, B and C, and pairing the wrong types across a trunk and a harness produces a link that is physically intact and logically reversed. Label both ends, record the polarity type on the trunk, and keep the total channel insertion loss comfortably inside the module’s budget rather than at its limit.
Breakout Cables and Port Economics
Ports, not bandwidth, are usually the scarce resource. A breakout assembly lets one high-speed port serve four or two slower ones: a 100G QSFP28 port splits into 4 x 25G for SFP28 servers, a 200G QSFP56 port splits into 2 x 100G, and a 400G port splits into 4 x 100G. On a 64-port 200G platform like the SN4600 VS2RC or a 32-port SN3700 VS2R, breakouts are how a 200G fabric gets connected to a fleet of 100G hosts without buying twice the switching.
Two rules keep breakouts out of trouble. The assembly must actually be a breakout type, not a straight cable with the wrong shell on one end, and the switch must support port splitting on that specific port group in that firmware release. Also decide the uplink oversubscription before, not after: 64 ports broken out to 4 x 25G gives 256 host links, and the spine now has to absorb four times the fan-in that the datasheet implied.
Qualification Traps That Cause “Link Down”
Most cabling failures are configuration and qualification failures, not hardware failures. A passive DAC that is not on the platform’s qualified list for that length and speed may still link, and may still pass traffic in a bench test, while sitting outside the signal margin the ASIC was tuned for. That shows up weeks later as CRC errors under load. Always check the length, the speed and the firmware release against the vendor’s cable compatibility matrix, and prefer assemblies whose EEPROM coding matches what the platform expects.
Beyond qualification, four settings cause most of the rest: FEC mode mismatched between ends, auto-negotiation disabled on one side and enabled on the other, a port forced to 40G when the cable and the far end are running 100G, and a speed mismatch inside a breakout group where one lane trains at a different rate than its siblings. When a link will not come up, work outward from the port configuration to the cable, and read per-lane error counters rather than trusting the link-up LED. A link that is up with rising symbol errors is worse than a link that is down, because it will silently cost you throughput for weeks.
Running the Numbers for One Rack
A workable method for a single rack: list every link with its measured length plus slack, bucket the links into under 3 m, 3 to 5 m, 5 to 30 m and over 30 m, then assign passive DAC, passive or active copper, AOC and optics respectively. Count the optical ports, multiply by module power, and check whether the switch and the rack cooling can carry the result. Add a 10 percent spare ratio per length for copper and per speed for optics, because a dead DAC means a dead port until a replacement arrives.
Then count part numbers, not just parts. One rack that uses eight different DAC lengths is a stocking problem, a labelling problem and a picking-error problem. Standardising on three lengths and accepting a little slack usually reduces total spend even when the per-cable figure is unchanged. If you are refreshing a rack and want the reach ceilings checked against the platforms you actually run, the switch and adapter listings on zhongxindata.hk carry the port types and speeds you need to match, and you can send a link list on WhatsApp to +86 13068616914 and we will flag the combinations that will not train before you order a single assembly.
Frequently Asked Questions
Is an AOC just two transceivers glued onto a fibre?
Functionally yes, mechanically no. An AOC contains an optical module at each end permanently terminated to a multimode ribbon and sealed at the factory, so it delivers the reach of optics with the handling of a cable. You cannot swap one end, shorten it, or clean the endfaces, and the length is fixed at order. In exchange you get lower total power than two discrete modules plus a patch cord, fewer part numbers and no field mating step to contaminate.
Can I push a passive DAC past 5 metres at 100G?
Not within specification. Passive copper for 100G QSFP28 is qualified to about 5 m, 200G QSFP56 to about 2.5 m, and 400G to 2.5 m on QSFP-DD or 3 m on OSFP. A longer passive assembly may link on a bench and then fail under thermal load or produce symbol errors, because the channel loss exceeds what the host equaliser can recover. If the run needs more, use active copper for a few extra metres or an AOC for anything beyond that.
Do DAC links need FEC enabled?
Often not at 100G, which is one reason copper links negotiate quickly and behave well on older firmware. 100GBASE-SR4 mandates Reed-Solomon FEC, and 200G and 400G PAM4 links essentially always run it. What matters is that both ends agree: a FEC mismatch is a classic “light but no link” fault, and it appears whenever one side is left on a default that the other side does not expect. Confirm the setting on both ends before replacing hardware.
Can I mix DAC and AOC links inside the same LAG?
It will usually form, but it is a poor idea. Members with different media have different latency, different FEC behaviour and different failure modes, so a bundle can end up with skewed traffic and unhelpful error counters when one member degrades. Keep each LAG or MLAG member set on the same media and the same length where possible, and reserve mixed media for links that are genuinely independent paths.
Can I reuse an existing OM3 plant for 100G?
Sometimes, with a shorter reach. 100G SR4 is specified to 100 m on OM4 but only about 70 m on OM3, so any run longer than that needs a different grade or a different topology. Also check the connector type and polarity: 100G SR4 uses eight active fibres in an MPO-12, and a legacy 12-fibre trunk wired with the wrong polarity type will look fine and pass nothing. Inspect and clean every endface before you blame the modules.
Request a quote for this configuration