AI Infra Interviews logo
← Hardware & Cluster Build-Out
Foundational

Cables, Transceivers and the Optics Power Budget

Cable selection depends on distance, port compatibility and power. Count both ends of optical links and check the exact part numbers before ordering.

TL;DR: Choose a cable or optical module that supports the required speed, reach and connector. Passive copper is useful over short distances; active cables and fiber extend the options. At cluster scale, transceiver power and cooling become substantial, so include them in the bill of materials and rack budget.

The ladder, by distance

TypeTypical reach at 800GPower per endWhere it is used
Passive DAC (direct attach copper)about 2 mnegligibleinside a rack, node to top-of-rack
Active copper (ACC or AEC)a few metresa few wattsadjacent racks, short intra-row runs
Active optical cable (AOC)tens of metressimilar to a transceiver pairrow to row, fixed length
Multimode transceiver (SR8) plus fiberroughly 50 to 100 m8 to 15 Wleaf to spine within a hall
Single-mode transceiver (DR8) plus fiberup to 500 m8 to 15 Wacross a hall, between halls, with breakout

The reach numbers come from vendor documentation rather than from a standard, and they move with signal-integrity work, so treat them as the current state of a moving target and check against the specific part number before ordering. The generation-to-generation shrink is the part worth memorizing: higher signaling rates can reduce passive-copper reach, so a previous generation’s cable choice needs to be checked again.

Choose a supported cable for the whole run

Distance narrows the choices. Speed, connectors, breakout support and the exact vendor part determine which choice works.

How far must the link travel?

Short run within a rack

  1. Check passive copper

    Confirm the supported reach at the required signaling rate. A cable that worked at a lower rate may not qualify.

A few meters between racks

  1. Check active copper

    Include the active electronics and their power in the endpoint budget.

Across a row or hall

  1. Check an optical option

    Compare fixed-length AOC with separate transceivers and fiber; verify reach, fiber type and connectors.

Longer path

  1. Check the single-mode configuration

    Verify the exact module, fiber and supported breakout combination.

  1. Count physical modules

    A logical link or breakout does not necessarily equal two separate optical modules.

  2. Assign power to the hosting equipment

    Put each module’s heat into its switch or server budget.

The table above provides representative ranges. Use the exact part specifications to approve a run; distance alone is not a compatibility test.

The power nobody budgets

Optical power depends on physical module count, not just the number of logical links. Breakouts, twin-port modules and copper links change that count. Use this hypothetical all-optical case to see the scale:

1,024 links x 2 separate modules per link = 2,048 modules
assumed module power = 15 W
optical power = 2,048 x 15 W = 30.72 kW
compared with 800 kW of compute load: 30.72 / 800 = 3.84%
sanity: replacing two logical ports with one twin-port module changes the module count.

Allocate that heat to the equipment hosting each module. For example, 72 modules at 15 W add 1.08 kW at a switch, beyond its other components. Check the supported module power and airflow for the specific switch.

Copper where you can, and why NVL72 is built that way

Short copper connections can avoid the optical conversion and modules required by a fiber path. That makes copper useful within a rack-scale NVLink system. Distance, signal integrity and service design still constrain the choice; it is not a reason to assume every scale-up fabric must stop at one rack.

The failure modes that trace back to cabling

A link that does not establish needs checks of port settings, module compatibility, polarity, reach and physical connections. A flapping link can point to signal quality, connector damage or temperature. Compare both ends’ counters and the supported configuration before replacing parts.

For a slow rail, verify the cable map as well as the link rate. Errors that appear only under load call for traffic and thermal measurements; they do not establish a thermal cause by themselves. NVLink and Fabric Faults describes how to narrow the affected path.

What interviewers are listening for

Explain how distance and compatibility narrow the cable choice. Then count the modules and their power, including both ends where separate optical modules are used. A useful answer also covers spares, supported breakout configurations and how link errors will be diagnosed.

Key takeaways

  • Check speed, reach, connectors and compatibility for the exact part number.
  • Count physical modules, including both ends where each uses a separate transceiver.
  • In the hypothetical example, 2,048 modules at 15 W consume 30.72 kW.
  • Include optical heat in the host switch or server’s cooling budget.
  • Link symptoms narrow the investigation; they do not uniquely identify a cable fault.
RELATED CONCEPTS
LESSONS THAT TEACH THIS
PRACTICE THIS IN REAL QUESTIONS