Why silicon borrows its laser and the four ways to bring one; external laser modules, wall-plug efficiency and comb sources; fibre attach and packaging; putting electronics next to photonics; and the link budget from laser to detector, priced in picojoules per bit, with what published links actually report.
Concepts used here, and where they are explained. Each links to a glossary entry: this series glossary, or the glossaries of Fourier Optics for Inference, Compute in Transit, LLM Inference Simulators and FHE Accelerator Simulators for concepts those series already explain.
Silicon has an indirect bandgap: the lowest state of its conduction band and the highest state of its valence band sit at different crystal momenta. An electron and a hole can only recombine and emit a photon with the help of a phonon to carry the momentum difference, so radiative recombination is slow and loses out to non-radiative paths. Silicon makes poor light emitters and no practical electrically pumped laser. III–V semiconductors (InP, GaAs and their alloys) have direct gaps and make the lasers of every optical link.
The consequence
A silicon photonic chip is a modulator-and-detector chip: the light comes from a III–V device, attached, bonded, grown or fibred in (slide 02).
The laser is often the least reliable and most temperature-sensitive part of the link, which pushes it away from hot electronics (slide 03).
The opportunity
Keeping the laser separate lets one source feed many modulators, or many wavelengths (slide 05).
Every dB lost between laser and detector is paid in laser power at the wall, divided by the laser's efficiency: the thread through the rest of this deck.
Reviews: Liang and Bowers, “Recent progress in lasers on silicon” (Nature Photon. 4, 511, 2010); Komljenovic et al., “Heterogeneous silicon photonic integrated circuits” (J. Lightwave Technol. 34(1), 20, 2016).
02
Four Ways to Bring the Laser
Approach
How
Strengths
Costs
Example
External source
A packaged laser elsewhere, fibred to the chip's input couplers
Replaceable, cool, mature lasers; one source can feed many channels
An extra fibre coupling (loss) and a polarisation-maintaining path
Lattice mismatch makes defects; quantum dots tolerate them better
Chen et al.'s InAs/GaAs quantum-dot laser on silicon (Nature Photon. 10, 307, 2016): 62.5 A/cm2 threshold, over 105 mW, operation to 120 °C, and an extrapolated mean time to failure above 100,158 h from 3,100 h of ageing
The early results above show the direction rather than the state of the art; they are chosen because their numbers can be checked against the papers. Komljenovic et al.'s review covers the heterogeneous platform in depth (J. Lightwave Technol. 34(1), 20, 2016).
The choice is as much about where the heat and the failures go as about coupling loss: an integrated laser shares the chip's temperature, and a failed one takes the chip with it.
03
External Lasers and the ELSFP
Why move the laser out. In co-packaged optics the optical engines sit beside a high-power switch ASIC; lasers dislike heat, and are the part most likely to fail. A front-panel laser module keeps them cool and lets a failed one be swapped without opening the package.
The OIF External Laser Small Form Factor Pluggable (ELSFP) implementation agreement, OIF-ELSFP-01.0 (August 2023), specifies such a module: a front-panel, hot-swappable laser source with a blind-mate multi-fibre optical connector at the rear for eye safety, and management over CMIS; a later revision, 02.0, followed (OIF implementation agreements).
The co-packaged module it feeds: OIF's 3.2T co-packaged module agreement specifies 8 × 400G (FR4 or DR4) optics behind a 32-lane CEI-112G-XSR host interface (OIF-Co-Packaging-3.2T-Module-01.0).
The price is an extra fibre-to-chip coupling for the laser light, and polarisation-maintaining fibre: decibels that appear in the budget of slide 08. The data-centre series takes co-packaged optics further.
04
Wall-Plug Efficiency
Wall-plug efficiency (WPE) is optical power out over electrical power in. It converts every optical decibel in a link into electrical watts: the laser's electrical power is Poptical/WPE, and each 3 dB of extra loss doubles it.
Figure
What it is
Source
Up to 13%
Measured WPE of a short-cavity III–V-on-silicon C-band DFB laser (200 µm) at 20 °C, coupling up to 2 × 6 mW into the silicon waveguide
Laser makers advertise higher values for high-power CW DFB chips intended as co-packaged-optics sources; treat these as the vendors' claims until measured in a system.
WPE falls with temperature, and drive and cooling electronics sit outside it: a thermo-electric cooler can cost more than the laser. The interactive on slide 09 lets you set WPE and see the laser's share of the energy per bit move.
05
Multi-Wavelength Sources
A WDM link needs one carrier per wavelength. Three ways to provide them:
Laser arrays
One DFB per wavelength, each held on its grid slot by its own temperature or current. Simple and powerful, but the count of lasers, and of things that fail, grows with the channel count.
Mode-locked comb lasers
One laser emits many equally spaced lines. Liu et al.'s quantum-dot mode-locked laser grown on silicon gave 58 lines within 3 dB at 20 GHz spacing and carried 4.1 Tbit/s as 64 channels of 32 GBd PAM4, with an external optical amplifier (Optica 6, 128, 2019).
Kerr microcombs
A CW pump in a high-Q microresonator generates a comb by four-wave mixing. Rizzo et al. drove a silicon photonic link with one: 32 wavelengths at 16 Gb/s each, 512 Gb/s, with bit error ratios below 10−12 (Nature Photon. 17, 781, 2023).
The catch is per-line power. A comb shares its output between its lines, and the outer lines are weaker. The budget is set by the weakest line, not the average, so comb flatness becomes a loss term.
Ring-based WDM transmitters (deck 02) pair naturally with combs: one bus, one comb, one ring per line.
06
Fibre Attach and Packaging
Alignment is the cost. A single-mode fibre has to sit on a coupler to a fraction of a micrometre, and stay there through curing, temperature cycles and years. Arrays of fibres in a V-groove block are aligned together, often actively (with light on) and then fixed with adhesive.
The coupler decides the package: grating couplers take fibres from above, anywhere on the die, and allow wafer-level test; edge couplers need the fibre at a prepared facet but give lower loss and broader bandwidth (deck 01, slide 06).
Polarisation and lasers add their own requirements: polarisation-maintaining fibre for the laser input, isolation against reflections, and a thermal path that keeps rings and lasers stable.
Carroll et al. describe photonic packaging as orders of magnitude more expensive than electronic packaging and perhaps the most significant bottleneck to silicon photonics in volume (Appl. Sci. 6, 426, 2016).
07
Putting Electronics Next to Photonics
Integration
How
Example
Monolithic, “zero-change”
Photonic devices drawn in an unmodified CMOS SOI process, beside the transistors
Sun et al.: a microprocessor with more than 70 million transistors and 850 photonic components on one 45 nm SOI CMOS chip, communicating directly by light (Nature 528, 534, 2015); Stojanović et al. on 45 nm and 32 nm SOI platforms (Opt. Express 26, 13106, 2018)
Monolithic, photonics-enabled CMOS
A CMOS process modified to include optimised photonic layers
GlobalFoundries' 45 nm 45CLO platform, Rakowski et al. (OFC 2020, paper T3H.3)
3-D
A separate electronic chip bonded face-to-face onto the photonic chip, with dense short connections
Daudlin et al.: 80 channels at 10 Gb/s, 800 Gb/s, at 120 fJ/bit (arXiv:2310.01615; Nature Photonics, 2025)
2.5-D and side by side
Electronic and photonic chips on an interposer or substrate, or wire-bonded
Common in pluggable modules and co-packaged engines
Why distance costs energy. Every millimetre of wire between driver and modulator, or between photodiode and amplifier, adds capacitance; a capacitive load costs about ¼CV2 per bit (deck 02, slide 05), and detector capacitance sets receiver noise (deck 02, slide 12). Flip-chip and 3-D bonds keep these femtofarads small.
Monolithic removes the interface entirely but takes the photonics in a transistor-optimised process, or the transistors in a photonics-optimised one; 3-D lets each chip use its own best process.
Work backwards from the receiver. Its sensitivity is an OMA (deck 02, slide 12). With extinction ratio r = P1/P0, the average power that delivers a given OMA is Pavg = OMA·(r + 1)/(2(r − 1)): a modest extinction ratio needs well over the OMA in average power.
Add every loss in dB to get the laser power; divide by the wall-plug efficiency to get electrical power. Fibre is the smallest term in a data centre: SMF-28 Ultra is specified at no more than 0.32 dB/km at 1310 nm and 0.18 dB/km at 1550 nm (Corning PI-1424).
PAM4 raises the required OMA by about 4.8 dB at the same noise (deck 02, slide 01); a shared laser split N ways costs 10·log10N dB per branch.
Margin covers ageing, temperature and manufacturing spread; standards write their budgets the same way, from transmitter and receiver specifications plus a channel allocation.
09
Interactive: Link Budget and Energy per Bit
A link computed from laser to detector and priced per bit. Every default here is illustrative, chosen to look like the two kinds of link described in this series, not taken from any product; change them to match a datasheet. Fibre couplers means chip-to-fibre and fibre-to-chip together; the laser can be one per lane or one source split (or a comb shared) across all lanes; SerDes + DSP covers serialisers, clocking and any DSP. Fibre attenuation is Corning's 0.32 dB/km at 1310 nm, PAM4 adds 10·log103 dB to the required OMA, and the slide 08 relation converts OMA to average power.
4
100
-10.0
5.0
1.5
4.0
0.0
3.0
0.5
3.0
15
1.00
1.00
3.0
0
2.0
Aggregate rate
—
Total optical loss
—
Laser output per lane
—
Total link power
—
Energy per bit
—
Limits: one direction of one link; no FEC overhead; laser power is the average power needed at the receiver multiplied up by the losses, ignoring the modulator's extra penalty at its operating point (set a lower extinction ratio to include it) and the weakest-line penalty of a comb. Electronics energies are lumped per bit. A link described in a paper or a datasheet may include a different set of these terms (slide 10).
10
What Published Links Report
Energy-per-bit figures are easy to quote and hard to compare: each counts a different set of terms. Read what is included before reading the number.
Result
Energy
What it covers
Source
3-D integrated link, 80 × 10 Gb/s
120 fJ/bit
The authors' headline figure for the whole 3-D bonded link; see the paper for what it includes
Modulators alone reach femtojoules; drivers, receivers, clocking, lasers and tuning bring whole links to tens or hundreds of femtojoules at best, and to picojoules where DSP and retimers are counted.
The interactive's presets show the same spread: where the laser, the tuning or the electronics dominate depends on the architecture more than on the modulator.
11
Thermal Tuning in the Budget
Rings must be held on their wavelengths (deck 02, slide 08), and heaters draw power whether or not data flows. Per bit, a heater costs its power divided by the lane's data rate.
Heater power per ring
At 10 Gb/s
At 25 Gb/s
At 50 Gb/s
0.5 mW
50 fJ/bit
20 fJ/bit
10 fJ/bit
2 mW
200 fJ/bit
80 fJ/bit
40 fJ/bit
5 mW
500 fJ/bit
200 fJ/bit
100 fJ/bit
Arithmetic: mW ÷ Gb/s = pJ/bit. Heater powers are illustrative; measured heaters need 12.7–24.77 mW for a full π (deck 02, slide 08), and a ring needs a fraction of that depending on its offset.
Tuning can exceed modulation by orders of magnitude. A 0.9 fJ/bit ring modulator (Timurdogan et al.) beside a 2 mW heater at 25 Gb/s spends nearly a hundred times more on staying tuned than on modulating.
Rizzo et al.'s design estimates for a comb-driven ring transmitter make the same point: about 173 fJ/bit for ring tuning against 12.5 fJ/bit for modulation, with about 180 fJ/bit for the receiver (arXiv:2109.10297; estimates, not measurements).
Remedies: faster lanes (the same heater over more bits), wider FSRs so offsets are smaller, better process control, athermal designs, thermal isolation that raises heater efficiency, and sharing heat between neighbouring rings.
12
How Bandwidth Scales
bandwidth = fibres × wavelengths per fibre × bits per symbol × symbol rate
Each axis has its own price. More fibres cost packaging and connectors (slide 06); more wavelengths cost lasers or combs, multiplexers and tuning (slide 05, slide 11); more bits per symbol cost sensitivity (PAM4's 4.8 dB); a faster symbol rate costs modulator bandwidth, electronics and DSP.
Pluggable Ethernet optics have mostly scaled symbol rate and bits per symbol on a few fibres or wavelengths; dense optical I/O proposals scale wavelengths, at modest rates per wavelength, to keep electronics simple and energy low.
Where this goes next: the data-centre series applies these devices to pluggables and linear-drive optics, co-packaged optics, optical circuit switching and AI cluster networks; the AI-processing series to computing with light.
13
What to Take Away
Silicon borrows its laser. External, attached, heterogeneous or monolithic: the choice is about heat, reliability and replaceability as much as coupling loss. Co-packaged optics leans towards external front-panel sources (OIF ELSFP).
Every decibel is paid at the wall, divided by a wall-plug efficiency of about 13% in the measurement and 14% in the comb analysis cited here; vendors claim higher.
Packaging, aligning fibres to sub-micrometre tolerance and keeping them there, is a large share of the cost.
Energy per bit is laser, driver, receiver, clocking and tuning: modulators reach femtojoules, whole links tens of femtojoules to picojoules. Always ask what a figure includes (slide 10; the interactive, slide 09).
Tuning can dominate a ring-based link; faster lanes and better process control are the main levers.
Next: this is the end of the foundation series. The data-centre series builds on it; the glossary is the reference both later series link to.