Cabling_
DAC vs AOC vs AEC: Choosing GPU Rack Interconnect Cables
This article details how deployment crews select and install DAC, AOC, and AEC cables for GPU rack scale-out networks on InfiniBand or Ethernet fabrics, separating those choices from the copper NVLink domain inside NVL72-class racks.
Key facts
- DAC cables carry electrical signals over twinax without retimers or optical conversion and remain the default for sub-rack and adjacent-rack links when loss budgets allow.
- AOC cables convert electrical signals to optical at each end and are used when reach exceeds passive copper limits or when weight and airflow matter more than power draw.
- AEC cables insert active retimers or equalizers inside the copper assembly to extend reach while keeping both ends electrical.
- Inside NVL72 racks the NVLink fabric runs exclusively over the copper backplane or spine; scale-out ports on the switches use separate QSFP or OSFP connectors for the InfiniBand or Ethernet fabric.
- Factory-terminated MPO or MPO-16 trunks carry parallel optics for AOC links; field crews perform cleaning, inspection, and polarity verification only.
- Power consumption at the port rises when moving from passive DAC to either AEC or AOC because active components draw from the host switch or NIC.
DAC limits and typical use inside the row
Direct-attach copper cables remain the lowest-power and lowest-latency choice when the link distance stays inside a single rack or between two adjacent racks. The passive twinax assembly introduces only conductor and connector loss, so crews verify the total insertion loss against the switch port budget before accepting the cable. Because no active components sit in the path, thermal load on the switch ASIC stays minimal and airflow paths stay clear.
Field crews pull DACs first during rack integration because they require no separate power budget calculation beyond the host port itself. When the measured loss exceeds the port specification, the run moves to an active solution rather than attempting to compensate with higher transmit power.
AOC selection when reach or weight drives the decision
Active optical cables convert the electrical signal to light inside the module at each end, allowing longer runs between rows or across the data hall without the bulk of thick copper. The optical engine adds power draw at both ends and requires clean fiber end-faces, so crews treat every AOC connection as an optical link that must be inspected and cleaned before mating. Polarity must be verified end-to-end because the factory-terminated MPO connectors fix the fiber mapping.
AOC weight and bend-radius advantages become decisive when cable trays above the racks are already dense. Crews still confirm that the added module power does not push any switch power-supply rail over its continuous rating before releasing the rack for commissioning.
AEC as the copper extension option
Active electrical cables place retimer or equalizer silicon inside the cable assembly to compensate for channel loss while keeping both ends electrical. This approach avoids optical conversion yet extends reach beyond passive DAC limits without changing the connector type at the switch. Power is still drawn from the host ports, but the cable itself becomes thicker and less flexible than a passive twinax run.
AEC is chosen when the design already standardizes on copper connectors for sparing and when the extra reach avoids the cost of an optical transceiver pair. Installation crews must still observe the same bend-radius rules as passive copper because the active components sit inside the jacket and can be damaged by sharp kinks.
Power, thermal, and airflow trade-offs across the three types
Passive DAC draws the least power and generates the least heat at the switch faceplate. Both AEC and AOC increase port power consumption because active silicon or optical engines sit inside the module housing. Crews calculate the cumulative draw across an entire switch line card before mixing cable types on the same power domain.
Airflow is affected when thicker AEC or AOC cables block more of the exhaust path than thin twinax. In liquid-cooled racks the effect is smaller, yet crews still dress cables to maintain front-to-rear or rear-to-front patterns specified by the rack OEM.
Common field failure modes and how crews catch them
The most frequent failure is dirty or damaged end-faces on AOC MPO connectors that produce high bit-error rates only after the fabric is loaded. Crews catch this by performing a one-sided inspection with a calibrated fiber microscope on every connector before mating and by re-cleaning any ferrule that shows debris or scratches. Another recurring issue is polarity reversal on MPO trunks that passes a simple continuity test yet fails link training; the remedy is to verify the full mapping with an MPO continuity tester before the trunk is dressed into the tray.
Over-tight bend radii on any copper assembly, whether DAC or AEC, cause intermittent errors that appear only after thermal cycling. Crews measure actual bend radii against the OEM minimum during routing and reject any run that violates the limit. Finally, mismatched cable lengths that exceed the supported reach for the chosen speed cause link flaps under load; crews confirm the length marking against the port configuration sheet before final dressing.
Installation sequence that prevents later rework
Crews complete all copper NVLink backplane connections first, then install the scale-out fabric cables. This order avoids disturbing finished NVLink assemblies when pulling longer AOC or AEC runs. After routing, every connector is inspected, cleaned if required, and mated only once.
Testing follows a strict order: first continuity and polarity on copper links, then optical inspection and power measurement on AOC links, and finally end-to-end BER or PRBS testing once the switches are powered. Any link that fails is isolated and replaced before the rack moves to liquid-cooling fill and system burn-in.
Standards referenced: IEEE 802.3 Ethernet specifications for 400 Gb/s and 800 Gb/s electrical and optical interfaces · InfiniBand Trade Association specifications for high-speed copper and optical links · TIA-568 series for structured cabling practices and connector inspection
Frequently asked_
Can a single cable type be used for both NVLink and the scale-out fabric?
No. NVLink inside NVL72 racks runs over the copper backplane or spine supplied by the rack OEM. Scale-out ports on the switches use separate DAC, AOC, or AEC assemblies for the InfiniBand or Ethernet fabric. Mixing the two domains creates immediate compatibility failures.
How do crews verify AOC polarity without a live switch port?
They use a dedicated MPO continuity tester that illuminates each fiber in sequence and confirms the expected mapping at the far end. This check occurs before the trunk is dressed into trays so that any reversal is corrected while the cable is still accessible.
What changes when moving from DAC to AEC on an existing switch?
The port must support the AEC retimer protocol and the power budget must be recalculated. Crews also confirm that the thicker AEC cable fits the existing cable management without violating bend radius before the change is accepted.
Why inspect every AOC connector even when it arrives factory-terminated?
Dust caps can fail or become contaminated during shipping and handling. A single dirty ferrule raises BER on that lane and can force the entire link to a lower speed. Inspection with a microscope before mating eliminates this source of later troubleshooting.
When does Leviathan Systems recommend AEC over AOC between rows?
When the design already standardizes on copper connectors for sparing and the added reach of the active retimer stays within the switch port budget. AEC avoids the need to qualify optical transceivers while still meeting the distance requirement.