The backbone of data centers, 5G networks, and long-haul telecom runs on optical communications — lasers carrying data as light through fiber. The interface between the fiber and the switch is the optical transceiver: a standardized, pluggable box that slots into switches, routers, and server NICs, converting electrical signals to light (transmit) and light back to electrical signals (receive).
From 10G to 100G, 400G, and the now-volume 800G, module rates double every few years — the core driver of data-center expansion. This article answers one question: what’s inside a modern optical transceiver? The answer: two chains — the transmitter (laser + driver + modulation) and the receiver (photodetector + TIA + DSP) — layered with two technical routes (IM-DD and coherent), SFP-to-OSFP packaging, and silicon-photonics integration.
1. Two chains: the transmitter and the receiver
A transceiver contains at least two parts:
Transmitter (Tx) — modulates the SerDes electrical signal onto a laser and sends it into the fiber. Typical components:
- Laser source: provides continuous wave (CW). Three mainstream types: DFB (distributed feedback, single-mode, narrow linewidth; the 1310/1550 nm data-center workhorse)[1]; EML (DFB laser + electro-absorption modulator EAM on one chip; the 25G/50G/100G-per-lane mainstream — fast, good extinction ratio)[3]; VCSEL (vertical cavity, 850 nm, low cost, arrayable; short-reach multimode under 100 m)[6]
- Driver: amplifies the SerDes’s low-swing signal into the voltage swing needed to drive the laser/modulator (the EML’s EAM needs about a 2–3 V swing)[3]
- Modulation: the EML uses electro-absorption (voltage changes absorption → changes output intensity); a DFB can be directly current-modulated, but at high speed the chir (frequency sweep) is large, so high-speed links use EMLs or an external Mach-Zehnder modulator (MZM)[3]
Key transmitter metrics: extinction ratio, chir, output power, linewidth — they set reach and bit-error rate.
Receiver (Rx) — reconstructs the electrical signal from the light. Typical components:
- Photodetector: a PIN diode (simple, low cost; the short/mid-reach mainstream) or an APD (avalanche photodiode, internal gain about 10–30×, high sensitivity for long reach/weak light, at higher noise and cost)[4]
- Transimpedance amplifier (TIA): the photodetector outputs a microamp-level current; the TIA converts it to voltage and amplifies it — the receiver’s most critical analog circuit, whose noise, bandwidth, and linearity set the receive sensitivity[5]
- CDR (clock and data recovery): recovers the clock from the data stream
- DSP: equalization, FEC, digital clock recovery (see section 4)
Key receiver metrics: sensitivity (minimum receivable optical power), saturation power, noise.
Figure 1 | Block diagram of a transceiver. Click for full size.
The module’s pluggable design is the key to the ecosystem: standard packages such as SFP and QSFP let modules from different vendors interchange, so switch vendors never have to build their own optical engines[1].
2. Two routes: IM-DD and coherent
Figure 2 | The two technical routes compared in modulation and detection. Click for full size.
IM-DD (Intensity Modulation - Direct Detection): the transmitter modulates only the optical intensity (NRZ or PAM4), and the receiver measures the intensity directly with a PIN/APD. Simple, low cost, low power — the mainstream for data-center short/mid reach (under 2 km); 10G/25G/50G/100G per lane all use it[1].
Coherent: the transmitter uses a narrow-linewidth laser + an I/Q modulator, modulating amplitude, phase, and polarization simultaneously (DP-QPSK, DP-16QAM); the receiver uses coherent detection (optical hybrid + balanced detection), mixing the signal with a local laser to recover the full amplitude/phase/polarization information. It carries far more data than IM-DD, and DSP can compensate dispersion — so it serves long reach/high capacity (long-haul telecom, DCI, 100G–800G per lane); the price is complexity, cost, and power[1].
In one line: IM-DD covers “near, cheap, low-power”; coherent covers “far, high-capacity” — IM-DD inside the building, coherent between buildings and across cities.
3. Packaging and silicon photonics
Figure 3 | Form factors and lane counts through the generations. Click for full size.
The form factor sets “how many lanes, what rate, what size”[1]:
- SFP / SFP+: 1 lane, 1G/10G
- QSFP+: 4 lanes, 4×10G = 40G
- QSFP28: 4×25G = 100G
- QSFP56: 4×50G = 200G
- QSFP-DD: 8 lanes (Double Density), 8×50G = 400G or 8×100G = 800G
- OSFP: 8 lanes, 400G/800G, larger with better thermal performance
The trend is “per-lane rate doubling + more lanes”: 10G→25G→50G→100G per lane, combined with 4/8 lanes to make 100G/400G/800G modules. QSFP-DD and OSFP are the two mainstream packages for 400G/800G today.
Conventional modules build their lasers, modulators, and detectors from discrete III-V semiconductor (InP, GaAs) components. Silicon photonics instead builds the optical components (waveguides, modulators, couplers, detectors) in a silicon CMOS process: low cost (mature CMOS volume manufacturing, sharing production lines with electronic CMOS), high integration (photonic integrated circuit, PIC), and co-packaging with silicon electronics (DSP, driver)[7]. The limit: silicon is indirect bandgap and cannot emit light — a silicon-photonics module still needs an external III-V laser (InP DFB/EML or VCSEL) as the source, coupled into the silicon waveguide. That is the standard architecture today: silicon PIC (passive + modulation + detection) + external laser. Silicon photonics is the mainstream direction for 400G/800G modules — integrating eight lanes of optics on one PIC cuts size, cost, and power[7].
4. The DSP’s role and the rate path
Modern high-speed modules (especially coherent and high-per-lane IM-DD) rely on the DSP for the critical processing[1]:
- Equalization: compensates the fiber/module bandwidth limits and dispersion (DFE, FFE)
- FEC (forward error correction): corrects transmission errors — the key that makes high-speed PAM4 work (e.g., RS-FEC)
- CDR: digital clock recovery
- Automatic gain/bias control: closed-loop control of the laser and receiver
The DSP is the module’s biggest power and cost item (especially the coherent DSP) — DSP evolution (more efficient FEC, lower power) directly sets the module’s energy efficiency.
Module total rate = per-lane rate × lane count; the current evolution path[1]:
| Generation | Per-lane | Lanes | Total | Modulation |
|---|---|---|---|---|
| 10G | 10G | 1 | 10G | NRZ |
| 100G | 25G | 4 | 100G | PAM4/NRZ |
| 400G | 50G | 8 | 400G | PAM4 |
| 800G | 100G | 8 | 800G | PAM4 |
The key is the per-lane rate: 100G per lane (PAM4) is the core of the 800G module. Every per-lane doubling doubles the bandwidth demand on the EML, TIA, and DSP — that is where the technical challenge lies. 800G (8×100G PAM4) is the current volume workhorse, widely used in AI clusters and large data centers.
Where they go: data centers are the largest market (intra-building under 100 m with VCSEL multimode, inter-rack under 2 km with SR/DR single-mode, inter-building DCI with coherent); 5G fronthaul/midhaul uses 25G/50G/100G modules; long-haul telecom uses coherent (100G–800G per lane); AI clusters use 400G/800G modules in the back-end network, scaling with GPU cluster size — the main driver of recent demand growth.
Back to the opening question: a modern optical transceiver contains two chains — the transmitter (laser + driver + modulation) and the receiver (photodetector + TIA + DSP) — layered with the IM-DD/coherent routes, SFP-to-OSFP packaging, and silicon-photonics integration. Scope: this article covers the current mainstream pluggable module specifications. Unresolved limits: pushing per-lane toward 200G needs silicon photonics and more efficient components, and power and cost remain the expansion bottlenecks — silicon-photonics volume manufacturing and DSP efficiency set the pace of the next generation.