Optical Module Datasheets: Six Numbers That Decide Whether a Link Works

Three quotations for the same part land on the same afternoon. 400GBASE-DR4, 500 m, MPO-12. Receiver sensitivity reads −5.9 dBm on the first, −4.4 dBm on the second, −3.9 dBm on the third. Sorted by that column, the first one wins, and the buyer saves a few dollars a port.

None of the three is lying. They are not measuring the same thing, and the 2 dB between them is a definition, not a capability.

A transceiver datasheet is a sales document with engineering units bolted on. Six numbers decide whether a link is still clean in year three: launch power, TDECQ, sensitivity at the right BER, overload, the wavelength window, case temperature. The rest is decoration, and some of it is a trap. Here are the six in the order they bite, plus the arithmetic that exposes a table assembled to flatter itself.

One receiver, five published numbers

The same 400G DR4 receiver can be described five ways, all compliant with the same clause of IEEE 802.3bs.

Put −5.9 and −3.9 side by side in a spreadsheet and you have invented 2 dB of margin that does not exist. That is the whole problem with comparing quotes by one column: the cheapest-looking module is often just the one whose vendor picked the friendliest definition.

My rule after too many of these: if a sensitivity figure arrives without a bit error rate, a temperature, and a qualification ("OMA", "average", "with SECQ"), it is not a specification. It is a number that happens to be on the page.

Launch power is the weakest number on the page

Most tables lead with Tx minimum and maximum average power, because those two cells are easy to test and easy to compare. They are also the least informative. IEEE 802.3bs marks the minimum launch-power row itself as informative — an explicit warning that power alone says little about signal quality.

What carries the signal is OMA — the optical modulation amplitude, the distance between the outer PAM4 levels that a receiver actually slices. For 400GBASE-DR4 the OMA outer window is −0.8 to +4.2 dBm, and the standard constrains it further with a combined limit: launch power in OMA minus TDECQ must stay above −2.2 dBm. Read that constraint twice. It says a transmitter may not buy eye closure with amplitude; whatever the equalizer recovers, the pair of numbers has to hold together.

Extinction ratio still appears at a 3.5 dB minimum, low next to the 6 dB that 10G NRZ parts carried. Deliberately so: PAM4 stacks three eyes inside one swing, where the ratio between the top and bottom level says far less about the middle eye than it did with two levels. ER survives as a check on laser bias, not as a quality metric.

If a datasheet gives you average launch power and nothing else, ask for OMA outer and TDECQ. A module can sit comfortably inside a +4 dBm average limit while its outer eye is closing, and the average-power column will never show it.

Sensitivity is a number without a subject until you attach a BER

Ten-gigabit parts are quoted at 10⁻¹². Four-hundred-gigabit parts are quoted at 2.4 × 10⁻⁴. Those two figures are six decibels apart in required signal-to-noise ratio — Q of 7.03 against Q of 3.49, or 16.9 dB against 10.9 dB of electrical SNR.

That explains a conversation I have every month. An engineer used to −15.5 dBm receivers sees a 400G DR4 quoted at −4.4 dBm and concludes that high-speed optics are fragile. They are not; the promise changed. At 400G the module no longer has to deliver error-free bits. It delivers a raw error rate the KP4 decoder can clean up — RS(544,514), thirty parity symbols per block, fifteen corrections, output at 10⁻¹². The 2.4 × 10⁻⁴ threshold is the decoder's appetite, not the receiver's limit.

A sensitivity figure therefore only means something once you know which side of the decoder it describes. Two vendors can both write "sensitivity −6 dBm, 400G" and mean receivers a full dB apart, because one tested at the FEC threshold with an ideal transmitter and the other with a stressed one.

TDECQ is the number that replaced the old eye mask. Nobody draws a polygon over a PAM4 eye any more; the eyes are closed before equalization, so every module would fail. Instead the transmitter is measured against a reference receiver with a real equalizer, noise is added until the symbol error rate hits 4.8 × 10⁻⁴ — twice the bit rate, because Gray coding turns most symbol errors into a single bit — and the added noise in dB is the penalty. The ceiling sits at 3.4 dB across most 400G/800G reaches and 3.9 dB on the long-reach variants; the 1.6T amendment tightens both.

A vendor shipping at 3.0 dB is not being generous. It spends the difference on the year a connector gets dirty and nobody notices.

The ceiling nobody checks

Every link budget I see is built from the bottom: launch power minus span loss must clear sensitivity. Almost none carry a top constraint, and 400G/800G links fail from the top more often than people expect.

400GBASE-DR4 tolerates +4.0 dBm of average received power, with a damage threshold of +5. The transmitter is allowed to launch +4.0. Put those two facts on a three-metre jumper with 0.5 dB of loss and a pair of APC endfaces that have been in a pocket, and the receiver is working within half a decibel of saturation. Link comes up. Errors appear at week six, in the afternoon, on the two ports with the shortest cables. The module gets replaced twice and the cable never does.

Reflectance makes it worse. The standard asks for −26 dB transmitter and receiver reflectance and allows 21.4 dB of return-loss tolerance. A scratched APC endface, or a UPC endface mated to an APC one, breaks that quietly: the penalty lands as jitter rather than loss, so the power reading stays healthy.

The fix costs almost nothing: on any 400G or 800G link shorter than about 20 m, check where the receive power sits against the overload column, not against sensitivity. If it is within 2 dB of the ceiling, add a fixed attenuator. Five dollars against a truck roll.

DDM telemetry hands you receive power and case temperature every five seconds, which makes the check trivial — provided the dashboard plots the overload line too.

Two columns pick the SKU: wavelength and case temperature

Wavelength gets read as a colour. It is a window, and the width matters.

400GBASE-FR4 and LR4 run four CWDM lanes on a 20 nm grid from 1271 to 1331 nm, each lane allowed ±6.5 nm of drift. 400GBASE-DR4 sits at 1311 nm with a 13 nm window, 1304.5 to 1317.5. That is why "1310 nm" is not a specification you can design a plant around: a DR4 lane and an FR4 lane are both "around 1310", and they are not interchangeable in the same mux. Ordering by description rather than by lane plan is how a 2 km span ends up with four modules that all light up and none that pass traffic.

Case temperature is the other quiet one. The commercial window is 0–70 °C on the case, not on the aisle, and a case runs well above inlet air once a 1U chassis holds 32 cages. Extended-range parts claim −40 to 85 °C; the honest question is whether the optical numbers there are the same numbers or a derated set measured at the corners. Ask for corner data. Identical launch power across a 125 °C span is a design claim, not a line item.

Here is what our own columns say — 10G SFP+, 100G QSFP28, 400G QSFP-DD, 800G OSFP — read the way a buyer should read them.

The DR4 row is the whole article in one line. Anyone can produce a number. Producing a number that survives subtraction against a row measured the same way is the harder discipline, and it is the one worth testing suppliers on.

None of this reaches a purchase order unless you ask for it: worst-case values guaranteed across the case temperature range and end of life, the quantity behind each sensitivity figure, OMA and TDECQ rather than average power alone, overload and damage thresholds, and a batch test report instead of one golden sample. A supplier who can answer all five in a single document is telling you something about the production line rather than the brochure — which, on an optical module datasheet, is the only interesting thing left to learn.

FAQ

Why do two vendors quote different receiver sensitivity for the same 400G DR4?

Because at least four quantities get called "sensitivity": average received power, OMA outer, stressed sensitivity with eye closure, and sometimes a room-temperature typical. On 400GBASE-DR4 those land at −5.9, −4.4 and −1.9 dBm in the standard's own tables. Compare like with like.

Does a less negative sensitivity number mean a worse module?

Not necessarily — it may mean a stricter definition. A module quoted at −3.9 dBm with a stressed condition can outperform one quoted at −5.9 dBm with none. The number compares only inside the same definition and BER target.

What TDECQ should I accept: the 3.4 dB ceiling or something tighter?

3.4 dB is the compliance ceiling for most 400G and 800G reaches, not a target. Large cloud buyers write 1.4 dB into their own specifications for a reason: margin at the transmitter is margin at commissioning, when the plant is dirtier than the drawing. For AI fabric ports that cannot afford a flap, 2 dB or better is a reasonable ask.

My 400G link shows +2 dBm receive power and still throws errors. How?

Probably near the overload ceiling rather than sensitivity. At 400G, +2 dBm can sit within 2 dB of a +4 dBm limit, and a marginal connector adds jitter that never shows as loss. Check overload, reflectance and endface condition before swapping modules.

Can I use a CWDM4 module where the design says 1310 nm?

Only if the lane plan matches. CWDM4 and LR4 transmit four lanes between 1271 and 1331 nm; DR4 transmits a single lane at 1304.5–1317.5 nm. They are different optical interfaces that both get described as "around 1310 nm", and mixing them across a mux or a breakout produces links that light up and fail.

Which numbers should I demand in an RFQ to compare suppliers fairly?

Worst-case launch power in OMA with the TDECQ limit stated, sensitivity in the same quantity at 2.4 × 10⁻⁴, stressed sensitivity, overload and damage thresholds, wavelength window per lane, and case temperature range. Add a batch test distribution for TDECQ. With those seven cells you can rebuild any supplier's budget yourself — which is also the fastest way to find the one whose table does not add up.