Receiver sensitivity

Receiver sensitivity is the answer to the only question the receiving end of a link budget asks: how little light is enough? It is defined as the lowest received power at which the receiver still meets the specified error ratio, and it is meaningful only with the target attached. Legacy specifications used a raw bit error ratio of $10^{-12}$; FEC-based links specify the pre-FEC threshold instead ($2.4 \times 10^{-4}$ for the RS(544,514) code), because the forward error correction turns anything below that into an effectively error-free stream.

The idealized floor comes from the noise of the receiver itself. A decision between two levels separated by OMA fails at the target ratio when the level separation is $2Q$ times the noise, so a PIN-TIA receiver with input-referred noise current $i_n$ (RMS, over the receiver bandwidth) and responsivity $R$ needs

with $Q \approx 7.0$ for the $10^{-12}$ target and $Q \approx 3.5$ at the FEC threshold. The formula makes the design levers visible: an avalanche photodiode buys sensitivity through internal gain until its excess noise catches up, a better TIA lowers $i_n$, and FEC lowers $Q$, which is worth about 3 dB of optical power and is the reason every high-speed link now runs against the FEC cliff rather than $10^{-12}$.

Datasheet sensitivities sit well above this floor, and the gap is bookkept as power penalties: the transmitter's TDECQ allowance, finite extinction ratio if the specification is average-power referenced, intersymbol interference in the receiver's own front end, and multipath interference. PAM4 makes the accounting harsher than NRZ in a fixed way: three stacked eyes divide the same outer swing, costing $10\log_{10} 3 = 4.8$ dB of eye separation before any implementation loss. Representative unamplified 100G-per-lane PAM4 receivers land in the vicinity of $-5$ dBm OMA at the FEC threshold, which is the class of number the link budget has to clear.

Two conventions prevent most specification misreadings. Sensitivity is quoted either in average power or in OMA; the two differ through the extinction ratio, and comparing one vendor's average-power number against another's OMA number misstates the link by a dB or more. And sensitivity is paired with an overload ceiling: the usable window between too little light and too much is the receiver's dynamic range, which short links can exit from the top.

References: Agrawal, Fiber-Optic Communication Systems (4th ed.), Ch. 4; IEEE Std 802.3, PMD clauses (sensitivity and stressed-sensitivity definitions). The Datacenter Link Budget Explorer treats sensitivity as the floor of the loss waterfall, and the PAM4 Eye & TDECQ Explorer shows the transmitter-side penalties that datasheet sensitivities must absorb.