Inside DAC, AOC and AEC cables

A cable with transceiver-shaped ends is, to the host, a transceiver: it presents an identity, may present diagnostics and obeys the same pins. What is inside differs completely — copper pairs and a tiny EEPROM board, or a pair of optical engines with a fibre ribbon between them, or retimer chips that make copper reach further. This page opens the three families and shows how their construction shows up in memory, in DDM and in the failures you see.

The three families

Family	Inside	Reach (25G / 100G / 400G per direction)	Power per end	Management
Passive DAC (direct attach copper)	twinaxial copper pairs, 26–30 AWG, a paddle-card PCB with an EEPROM in each plug	3–5 m / 3–5 m / 2–3 m	~0 W (EEPROM only)	identity only, no DDM
ACC (active copper cable)	as DAC plus linear amplifiers/equalisers in the plugs	up to 5–7 m at 25–50G	0.5–1 W	identity; sometimes temperature
AEC (active electrical cable)	copper plus retimers/DSP in each plug — the signal is regenerated	3–7 m at 400G/800G	3–6 W	CMIS with lane states, often DDM-like monitors
AOC (active optical cable)	an optical engine in each plug (VCSEL array, PD array, driver, TIA, controller), MMF ribbon between	30–100 m, some 300 m	1–4 W	full DDM at each end (SFF-8472/8636/CMIS)

Host-side behaviour, auto-negotiation and coding conventions: Copper Ethernet & DAC; field faults: DAC & AOC problems.

Passive DAC

 plug A ─ paddle card (EEPROM, ground, pull-ups) ─ twinax pairs ×2 (SFP) / ×8 (QSFP) / ×16 (QSFP-DD) ─ paddle card ─ plug B

Electrical only: the host SerDes drives the copper directly; loss grows with length and frequency (≈ 10–20 dB at 12.9 GHz for 3 m of 30 AWG), which is why reach shrinks with each generation and why 25G+ defines cable classes by loss (CA-25G-N/S/L) and pairs them with FEC (FEC & link training).
Identity: an EEPROM (A0h only) in each plug — often different content per end (serial suffix, sometimes a different vendor if the cable was re-labelled). Key bytes: connector 23h, cable technology bits (passive copper), length in metres (SFP byte 18, QSFP byte 146, CMIS 202), attenuation at 2.5/5/7/12.9 GHz in the wavelength fields (Compliance codes).
No A2h, no temperature, no LOS — the host has nothing to monitor; presence and identity are all.
Failures: broken pair from bending (one lane dead), cracked paddle-card solder (intermittent), mismatched or missing identity (host rejects one end), too long for the host's equaliser (works in one switch, not another).

Active copper: ACC and AEC

ACC adds analog amplification to extend reach a metre or two; AEC adds a retimer or DSP per lane in each plug, so each copper run only has to cover plug-to-plug and the signal is regenerated — the way 400G/800G reaches 5–7 m over copper in AI racks. AEC plugs run warm (3–6 W), present as CMIS modules with data-path states and firmware, and draw real power from the port — check the power class as with optics (Power & thermal). Both are coded as active cable with the relevant SFF-8024 codes (BER-class AOC/ACC codes 01h/08h/18h/19h, 30h–33h) rather than as passive copper.

Active optical cable

 plug A: driver ─► VCSEL array ─► lens ─► MMF ribbon (OM3, 4/8/12 fibres) ─► lens ─► PD array ─► TIA ─► plug B
         TIA ◄── PD array ◄── lens ◄── ribbon ◄────────────────────────────── lens ◄── VCSEL array ◄── driver
         controller (DDM, identity)                                                controller (DDM, identity)

Each plug is a COB optical engine — bare VCSEL and PD dies under a moulded lens array, glued to the fibre ribbon; no connectors anywhere (Chip-on-board).
850 nm VCSELs on OM3/OM4 ribbon: cheap, low power, 100 m class; the wavelength and fibre are fixed, so an AOC is coded like an SR module (850 nm, MMF lengths) but with connector 23h.
DDM at each end is real: temperature, Vcc, per-lane bias, Tx and Rx power — the Rx numbers describe the internal link. A low Rx on one lane inside an AOC means that lane's VCSEL or fibre is failing; there is no connector to clean.
Breakout AOC (QSFP to 4 × SFP): the QSFP end carries one identity and 4-lane DDM; each SFP end its own single-lane identity and DDM.
Failures: bend radius violated (one or more lanes drop), engine overheating in dense faceplates, VCSEL ageing on one lane, a plug's controller dead (that end invisible while the other reads fine).

Coding the two ends

Hosts read each end independently; a cable is accepted only if both ends pass the policy of their respective hosts. Typical requirements:

Field	Rule
Identifier, connector, cable technology	identical on both ends; connector 23h
Vendor name / PN	identical; the PN usually encodes length and gauge
Serial number	usually the same base with an end suffix, or identical
Length byte	actual length in metres, same on both ends
Attenuation bytes (copper)	measured values for the assembly
Compliance / extended code	matching the rate class (e.g. 0Bh–0Dh for 25G CR classes, 40h for 50G/100G/200G CR)
Checksums	recalculated per end — Checksums

A cable that links from switch A to switch B but not the other way round usually has one end coded differently. Reading both plugs on the bench is the fastest check (Symptom index).

Why cables instead of optics

Criterion	DAC	AOC	Optics + fibre
Cost per link (short)	lowest	medium	highest
Power	none	low	low–high
Latency	lowest	+ ~few ns	+ ~few ns
Reach	metres	up to 100 m	km
Weight, bend radius, airflow blockage	worst	good	best
Field repair	replace whole cable	replace whole cable	clean or replace one part
Diagnostics	none	full DDM (internal)	full DDM (the real link)

In CodingBox

CodingBox reads each plug as the module it claims to be: DAC identities with length, attenuation and cable-class codes; AOC ends with full per-lane DDM; AEC ends as CMIS modules. Coding a cable means coding both ends consistently — the code database keeps a pair together, and the EEPROM editor handles the length and attenuation fields that optics do not have.

Cable reach classes and split configurations per InfiniBand generation: IB cabling; how the media mix plays out in a GPU cluster: Cabling an AI cluster.