800G DAC vs AEC vs AOC: In-Rack Cable Selection Table

Compare 800G DAC vs AEC vs AOC and ACC for in-rack links: reach, power, FEC margin, bend radius, and the vendor tests to demand before you qualify.

The 30-second answer: which cable for which in-rack distance

For 0.5–3 m links inside a rack, shortlist passive DAC first. Step up to ACC only when your measured route exceeds the DAC insertion-loss budget, to retimed AEC for row-to-row copper, and to AOC once the link leaves the row. The rule that survives qualification: choose the shortest, simplest cable that still passes your pre-FEC BER target at thermal worst case — not at bench temperature.

That is the whole decision at a high level. What follows is how to defend it in a BOM review, because at 800G the choice is no longer "short cable or long cable". It is a three-way trade between reach, power per end and FEC margin, and the published numbers you will be quoted do not always agree with each other.

Two caveats before any table: reach figures vary by vendor, by port type (OSFP versus QSFP-DD800), by wire gauge and by whether the link runs at 800G or 1.6T. Treat every number below as a datasheet item to be confirmed in writing, not a specification you can copy into a design document.

Master comparison table: DAC vs ACC vs AEC vs AOC at 800G

Read this as a reach ladder as much as a table — each row is a step up in signal conditioning, and each step buys distance with power, heat and complexity.

Cable familyTypical reach at 800GPower per endCable OD and bend radiusAirflow and cable managementCost directionMain qualification risk
Passive DAC (twinax)Shortest in-rack runs; commonly quoted up to ~3 m — verify per datasheetNo active power draw at the cageLargest OD and stiffest of the four; highest minimum bend radius — verify per datasheetBulky bundles; easiest to service, least forgiving of tight routingLowestMechanical, not electrical: kinks and over-tight bends after cable management
ACC (analog redriver / CTLE)Roughly 4–5 m at 800G in published comparisons — verify per datasheetLow; one published comparison lists ~1.5 W per end — verify per datasheetClose to DACSimilar to DACLow to moderatePaying for an active part that may add little margin over DAC at 800G
AEC (digital retimer with CDR)~6–7 m at 800G per MapYourTech (11 Jul 2026); some vendor guides claim up to 10 m — verify per datasheetPublished figures conflict: up to 5.5 W per port in one source, ~10–11 W per end in another — see the note belowModerate; active ends add bulk at the cageBetter than DAC at longer routes; still copper-stiffModeratePower budget and thermal headroom at the cage
AOC (electrical-optical-electrical)The only option once the link leaves the row; vendor guides cite the 100 m class — verify per datasheetModerate; confirm per datasheet, per endSmallest OD, lightest, most flexibleBest for dense tray fill; fixed length complicates sparesHighest at short reachConversion latency, fixed lengths, transceiver-grade failure modes at both ends

Why the published AEC power numbers disagree

One widely circulated comparison states that 800G AECs can draw up to 5.5 W per port (UByteLink, publisher metadata undated). Another puts 800G AEC assemblies at roughly 10–11 W per end based on vendor datasheets (MapYourTech, 11 Jul 2026). Both can be true at once, because "per port", "per end" and "per cable" are three different denominators, and a cable has two ends.

This matters more than it looks. On a 2,000-link cluster, an error of 5 W per link is 10 kW of continuous load before cooling is counted — and cooling is where the second-order cost sits. Treat any power figure you are given as incomplete until the vendor states the denominator, the data rate, the temperature and whether it includes the retimer only or the whole assembly.

On latency

Passive DAC adds propagation delay only. Active families add equalisation or retiming, and optical adds conversion at both ends; UByteLink's published comparison places DAC in the single-digit-nanosecond range and the active families two to three orders of magnitude higher. For most AI back-end fabrics this is not the deciding factor — FEC margin and thermal behaviour usually are. Confirm the delay budget with your workload owner rather than assuming it is negligible.

The four specs that actually decide an 800G link

1. Reach is an insertion-loss budget, not a tape measure

A "3 m DAC" is a statement about loss at a given gauge, not about how far the cable can physically be routed. Bend radius, cable ties, tray routing, connector fan-out and the number of mating cycles all consume margin before the cable itself does. Measure the actual route — including the service loop you will add later — and ask the vendor for insertion loss at the gauge you are buying.

2. Power per end and the cooling tax

The arithmetic is simple enough to reproduce in a spreadsheet: links × watts per end × 2 = continuous electrical load, then apply your own PUE. UByteLink's piece also proposes a cooling multiplier of 0.5–1.0 W of cooling energy per watt of link power; treat that as a publisher estimate with no stated method, useful for sensitivity analysis, not for a business case.

3. BER and FEC margin — the spec nobody puts in the quote

A link that trains on the bench can still flap in production once FEC-corrected error counters start climbing under thermal load. Ask for pre-FEC BER and corrected-error counts across your cold-aisle and hot-aisle temperature range, not just at room temperature. A cable that reports "link up" tells you nothing about how much margin is left when the rack is at full power.

4. Bend radius and cable OD are airflow decisions

In a dense rack the cable bundle is an airflow object. Smaller OD and lighter weight change how much tray you fill and how much cold aisle you block; passive twinax fights back in tight bends and behind PDUs. Make minimum bend radius, OD with tolerance and maximum tray fill explicit procurement requirements — otherwise you will discover them during cable management, after the GPUs are racked.

Priority order for the shortlist: FEC margin at thermal worst case first, then reach against measured route, then power per end against rack thermal headroom, then OD and bend radius against the physical layout, then cost. For scale-out leaf-spine links the order holds; for scale-up backplane links inside a chassis, power and OD usually outrank reach.

Why 800G and 1.6T shift the trade-off versus 400G

Per-lane signalling rates rise with each generation, so the passive copper reach that was comfortable at 400G shrinks at 800G and again at 1.6T. A rack layout that qualified under a 400G BOM is not evidence that it qualifies now — reach, wire gauge, connector type and FEC budget all need re-qualification.

Port count per rack also rises with GPU density, so a small power-per-end error multiplies into rack-level kilowatts. The counterweight is that the shortest, highest-count links stay copper. NVIDIA's published GB200 NVL72 rack design connects 72 Blackwell GPUs through a passive copper NVLink backplane — over 5,000 copper cable connections across four cartridge assemblies — and NVIDIA states that using copper there instead of retimers and optics avoids roughly 20 kW of additional draw per rack (as reported by MapYourTech, 11 Jul 2026; a vendor claim, not an independent measurement).

That is the pattern to copy into your own design: passive or lightly conditioned copper for the shortest, highest-count links; AEC for the copper stretches; AOC once the link leaves the row. Vitex's selection guide makes the commissioning argument plainly — in a 1,024-GPU cluster, server-to-leaf links number in the thousands, and specifying per link type from measured topology distances is what separates deployments that commission on schedule from those that replace a full cable set after installation (Vitex, publisher metadata undated).

Where 800G links fail in the rack: five failure modes

  1. Trains on the bench, flaps in production. Usually insufficient FEC margin after real routing shortened the effective budget. The cable that removes this risk is the one with a documented pre-FEC BER headroom; ACC and AEC may or may not move it — only a test report tells you.
  2. Thermal throttling or port shutdown at the cage. Active ends add heat exactly where airflow is worst. Passive DAC removes the risk entirely at the cage; AEC and AOC move it into your thermal budget, which is why the power denominator matters.
  3. Bend-radius violation during cable management. A kinked twinax behind a PDU is a dead link. Specify minimum bend radius and tray fill as requirements; AOC's smaller OD reduces the risk but does not eliminate it.
  4. Airflow and serviceability from bundle OD. Cold-aisle blockage and the inability to pull one link without disturbing its neighbours. Optical changes this materially; copper does not.
  5. Wrong reach specified, discovered post-rack. The expensive case is replacing a whole cable set after GPUs are installed. Measure the route before the PO, not after commissioning.

The qualification pack: what to ask your vendor to prove

This is the list worth attaching to an RFQ. It is the difference between a datasheet and evidence.

  • A test report at the actual data rate and the actual port type (OSFP or QSFP-DD800) — not a generic 400G report reused for an 800G part.
  • Pre-FEC BER and FEC-corrected error counts across your operating temperature range, not a single 25 °C point.
  • Insertion loss and reach data at the wire gauge you are buying, with a stated minimum bend radius and cable OD including tolerance.
  • An interoperability matrix naming the switch ASIC and port/module combinations tested, and confirmation that the cable is coded for your platform.
  • The commercial terms in writing: what warranty and return process covers a link that passes bench test and fails in-rack.
  • Whether samples are production-representative or hand-built, and at what volume the quoted lead time applies.

Shortlist rule per reach, port count and environment

  • Under 3 m, high port count, air-cooled: passive DAC first for lowest power, lowest added latency and the simplest failure mode. Verify insertion loss for the actual route.
  • 3–5 m adjacent rack with a marginal budget: ACC only if the vendor's datasheet supports the reach at your temperature; otherwise step straight to AEC rather than gambling on an intermediate step.
  • 5–7 m and beyond, copper only: AEC with retimer. Budget power per end at the datasheet figure and confirm the FEC and delay budget with the workload owner.
  • Off-row, high-EMI or routing-constrained: AOC, accepting conversion latency and fixed-length spares planning.
  • Most clusters are mixed. Three cable families in one rack is normal. The deliverable is a per-link-type allocation based on measured distances, not one cable type for the whole BOM.

Frequently asked

Is passive DAC lower latency than AEC or AOC at 800G?

Yes — passive DAC adds propagation delay only, with no equalisation, retiming or optical conversion stage. The absolute difference is small relative to most fabric latency budgets, but confirm it against your workload rather than assuming.

How far can an 800G DAC actually reach in a real rack?

Commonly quoted up to about 3 m, but the usable figure depends on wire gauge, port type, bend radius and how much FEC margin you are willing to spend. Measure the route including service loops and ask for insertion loss at your gauge.

Why do vendors quote different power figures for the same 800G AEC?

Because "per port", "per end" and "per cable" are different denominators, and some figures cover the retimer while others cover the whole assembly. Always ask which one applies, at what data rate and temperature.

Can I use a passive DAC for 1.6T in-rack links?

Only against a datasheet that states 1.6T reach for your specific port type and gauge — do not carry a 1.6T reach figure across from an 800G part or copy a 400G-era BOM forward. Re-qualify the link on your hardware.

Do AEC and AOC require FEC, and does that change my BER budget?

Retiming and optical conversion change the error profile but do not remove the need for a pre-FEC BER margin check. Ask for corrected-error counts under thermal load for every family you shortlist, DAC included.

Next step: send your port count, reach and rack layout

Send your port count, per-link measured reach, port type (OSFP or QSFP-DD800) and rack layout to request an 800G/1.6T cable sample and quote. What comes back is a per-link-type recommendation, the datasheet values for the specific gauge and length you are buying, and the qualification documents listed above — returned with the sample rather than promised in a brochure.

Be explicit about the honest limitation of this page: every reach and power figure quoted here is a vendor or third-party published value, not our own independent measurement. The two sources cited disagree on AEC power, and neither presents full test conditions. Final reach, FEC margin and thermal behaviour must be confirmed on your hardware, in your rack, at your ambient temperature — which is exactly what the sample is for.

Request a quote

Share a model or describe your application. Include a drawing link if you have one.