Cabling and Optics for AI Fabrics: DAC, AOC, 800G Transceivers and Link Budgets

Overview

The physical layer is the least glamorous part of an AI fabric and the most common source of intermittent cluster degradation. An 8-GPU node at 800G per GPU needs eight fabric links; a 512-GPU cluster runs thousands of cables and transceivers, and a single marginal optic among them can flap under load, trigger priority flow control storms and stall a training job that spans every node. Choosing correctly between DAC copper, active cables and pluggable optics – and enforcing link-budget margin, validation soak tests and labelling discipline – is cheaper than any amount of post-incident debugging.

The interconnect ladder: DAC, AEC, AOC, transceivers

Four media classes cover an AI fabric. Passive DAC (direct-attach copper) is a twinax cable with no electronics: near-zero power, near-zero added latency, lowest cost – but at 800G the loss budget caps it around 1.5-2 m, confining it to intra-rack links such as GPU node to top-of-rack or rail leaf in adjacent racks. AEC (active electrical cable) adds retimer chips to stretch copper to roughly 3-7 m at a few watts. AOC (active optical cable) fixes transceivers to both ends of a fibre, reaching 30 m or more; it is cheaper than pluggables per link but unserviceable – a failed end means replacing the whole assembly. Pluggable transceivers over structured fibre (typically OSFP modules at 800G, in DR8, 2xDR4, FR4-class variants for 100 m to 2 km) cost the most and burn 13-18 W each, but are field-swappable and reusable across generations of cable plant. The pattern that works: copper inside the rack, fibre between racks, pluggables wherever a link must be serviceable.

800G optics in 2026: DSP, LPO and CPO

Conventional 800G modules carry a DSP retimer that cleans up the electrical signal – robust, interoperable, power-hungry. Linear pluggable optics (LPO) remove the DSP and lean on the switch SerDes, roughly halving module power to around 10 W and cutting latency, at the price of tighter host-and-module matching. Co-packaged optics (CPO) move the optical engine onto the switch package itself; vendor roadmaps – including Quantum-X and Spectrum-X Photonics systems slated through the second half of 2026 – claim large power and reliability gains, as surveyed by MapYourTech’s CPO status review. For clusters bought in 2026, DSP pluggables remain the safe default, LPO is worth qualifying for power-constrained rows, and CPO is a switch-purchase decision, not a cabling one. Fabric-level context is in AI Fabric Fundamentals.

Link budgets: where the margin goes

Every optical link has a budget: transmitter launch power minus receiver sensitivity, typically a handful of dB for short-reach 800G parts. Fibre itself consumes little over data-centre distances; connectors and splices consume the rest, at roughly 0.2-0.5 dB per mated pair – more when end-faces are dirty. A patch-panel-heavy path with four connector pairs can eat half the budget before the fibre matters. The budget interacts with FEC: 800G links run at raw bit-error rates that only forward error correction makes usable, and the margin between operating BER and the FEC correction ceiling is the real health metric. Design rules that hold up in practice: keep at least 1.5-2 dB of unallocated margin, minimise mated pairs per path, and treat every connector insertion as a cleaning event – inspect-and-clean is the single highest-value habit in optical operations.

Why marginal optics degrade clusters intermittently

A marginal optic passes acceptance testing at idle and fails only under load. Full-power training raises ambient and supply stress; laser bias and receiver characteristics drift; pre-FEC BER climbs past the correction ceiling and the link flaps for a few seconds. In an RDMA fabric that is not a local event: PFC pause frames propagate, congestion spreads to healthy links, NCCL collectives time out, and a job spanning the whole cluster stalls or restarts from checkpoint. Because the trigger is thermal and load-dependent, the symptom appears as unexplained periodic slowdowns rather than a hard fault – among the most expensive failure modes to chase, as optical vendors’ AI-training analyses describe. The defence is telemetry: FEC corrected-codeword counters, DDM optical power and temperature per module. A rising FEC correction trend on one link predicts failure days in advance; swapping that module in a maintenance window costs minutes, not training days. This telemetry belongs in the same day-2 stack as Day-2 Operations for Rack-Scale AI.

Labelling and validation discipline

Large fabrics are debuggable only if the physical layer is documented. Minimum discipline: every cable labelled at both ends with a scheme encoding rack, node, NIC index and leaf port (rail designs depend on NIC k landing on leaf k – see the rail-optimised topology deep dive); a machine-readable cable database reconciled against LLDP after any change; polarity and end-face inspection on every fibre link at install; and a 24-72 hour burn-in soak under synthetic RDMA load with zero-flap acceptance criteria before a fabric carries production training. Hold 1-2 percent of each transceiver and cable type as on-site spares – relevant in India, where replacement optics can carry multi-week import lead times, and where higher ambient temperatures and dust make end-face contamination and thermal margin genuinely tighter than data-sheet conditions assume.

Media options at 800G compared

Frequently asked questions