Transceivers in AI clusters

AI training clusters are the largest consumers of high-speed optics ever built, and they push modules to their limits in rate, power and count. The engineering that matters is less about any single link and more about running tens of thousands of them reliably.

Which optics

Role	Module	Typical PMD	Connector
GPU server ↔ leaf (scale-out, IB NDR)	OSFP twin-port 2×400G	DR4 ×2	2 × MPO-12 APC
GPU server ↔ leaf (Ethernet)	QSFP-DD / OSFP 400G–800G	DR4, DR8, 2×FR4	MPO-12/16 APC, dual LC
Leaf ↔ spine	800G OSFP / QSFP-DD800	DR8, 2×FR4	MPO-16 APC, dual LC
In-rack short hops	DAC / AOC where reach allows	—	—

Single-mode DR optics dominate because they reach 500 m — enough for any hall — and scale from 400G to 800G on the same fibre plant. Angled APC connectors are required.

Power and heat

An 800G module draws on the order of 14–18 W; a 64-port switch faceplate holds a kilowatt of optics. OSFP's integrated heat sink is why it is preferred for the hottest modules; QSFP-DD relies on host cooling. Thermal throttling — a module backing off when it overheats — is a real cause of intermittent 400G/800G flaps (CMIS issues).

Linear-drive pluggable optics (LPO) remove the module's DSP to cut power and latency, moving equalisation to the switch ASIC; they are appearing on short DR links where the host supports them.

Reliability at scale

The arithmetic changes everything: with 100 000 optics, an annual failure rate of even 0.5% means ten replacements a week. Consequences:

Per-lane DDM monitoring across the fleet, trended, to replace a degrading module

before it stalls a training step (Failures).

Spares by batch and incoming inspection — failures cluster by production lot.
MPO cleanliness discipline — 16 fibres per connector, and one particle takes a

lane down (Link flapping).

A single slow lane slows the entire collective: tail latency is the metric.

Identity and interoperability

Large clusters standardise on a short list of qualified modules, and switch platforms validate identity. Third-party 400G/800G optics are used where the platform tolerates them and are coded to the expected identity where it does not — the same vendor lock rules as elsewhere, at far higher unit cost.

What is next

1.6T ports (OSFP-XD, 16 lanes) and 200G-per-lane electrical interfaces are the next step; co-packaged optics move the transceiver into the switch package for the largest systems.

In CodingBox

CodingBox reads 400G/800G OSFP and QSFP-DD modules through CMIS: module state, application advertisement, identity and per-lane Tx/Rx/bias on the DDM screen — the data an operator needs to qualify incoming batches, baseline new links and identify a single failing lane out of eight.