By the late 2000s the logic seemed inescapable. Multicore was scaling toward hundreds of cores, the electrical network-on-chip connecting them was projected to eat a growing share of the chip's power, and photonics had just demonstrated microring modulators a few microns across. If light beats copper across a datacenter, surely it should beat copper across a die? Thus was born the optical network-on-chip (ONoC) research programme — a decade of gorgeous architectures (HP's Corona, Columbia's photonic torus, Firefly, and dozens more) proposing to route cache lines between cores as wavelengths on waveguides. Corona's headline sketch: a 256-core chip wrapped in a photonic crossbar, delivering petabit-scale on-chip bandwidth at a fraction of the electrical power.

This lesson teaches the ONoC in two movements: first the genuinely elegant ideas — ring-based routers and wavelength-routed crossbars, which resurface across photonics — and then the sober accounting of why, fifteen years later, no shipping processor contains one. The autopsy is not a humiliation; it is the best masterclass this course has in where photonics wins, and it ends with the survivor: the part of the dream that is shipping.

The elegant machinery

The ONoC's element is the microring resonator from Module 3: a wavelength-selective tap that, when its resonance matches a passing channel, peels that one colour off the shared waveguide onto a drop port — and when detuned, lets it glide past untouched. A cluster of rings is therefore a router without a router: no arbitration logic, no crossbar transistors, just resonances deciding who hears what. Architectures assembled this into wavelength-routed crossbars: give every node its own wavelength, and a message's colour is its address. In a single-writer, multiple-reader (SWMR) design, each sender owns a wavelength on a waveguide snaking past every other node; receivers park a ring tuned to each sender they may hear from. The topology is a passive physical object — routing happens in the spectrum, contention is resolved by wavelength assignment before the fabric is even fabricated, and (echoing the previous lesson) nothing ever needs to be buffered mid-flight because nothing ever collides. On paper it is the most beautiful interconnect ever designed.

The sober accounting

Now the audit. Three line items, none fatal alone, jointly damning:

Laser power. The light comes from an off-chip laser with a wall-plug efficiency of perhaps 10–20%: delivering 10 mW of useful optical power costs 50–100 mW at the socket — before the ~5–10 dB of coupling and distribution loss between laser and the farthest detector, which multiplies the demand again. Worse, the laser burns this continuously, whether or not anyone is talking.

Thermal tuning. A ring's resonance drifts roughly 10 GHz per kelvin, and a many-core die is a weather system of hotspots. Every ring therefore carries a heater and a control loop pinning it on-station — of order a milliwatt per ring — and a wavelength-routed crossbar needs thousands of rings. The tuning bill alone can reach watts, also charged at idle.

The distances are too short. The crossover chart of lesson one cuts both ways. An on-chip hop is millimetres: at ~100 fJ/bit·mm, an electrical 5 mm hop costs ~0.5 pJ/bit, below the ~1–2 pJ/bit toll of an optical conversion. On-die, copper sits comfortably on the winning side of the crossover — and unlike the photonic fabric, its energy is activity-proportional: an idle electrical NoC costs almost nothing, while the photonic one pays its laser and heaters around the clock.

Run the totals: 4,000 rings × 2 mW + 2 W of laser overhead = 10 W of standing power, against an electrical NoC drawing a few watts only when busy. The beautiful crossbar loses the energy argument at precisely the distances a chip contains — and adds a control problem (per-ring calibration against process and thermal variation) that Module 11 will show haunts every large photonic system.

What survived

Strip the ONoC of its weakest claim — photons between cores — and a strong one remains: photons at the package boundary. The distances there are centimetres-to-metres (the right side of the crossover), the pipes are kept full (I/O traffic aggregates thousands of cores' demands), and the ring machinery is exactly right for dense WDM. That surviving idea is shipping today as in-package optical I/O — photonic chiplets flanking a processor die, a few pJ/bit, terabits per fibre — the direct heir of co-packaged optics, aimed at GPUs rather than switches. The honest division of labour that emerges from the ONoC's failure is this module's central theorem: electrons compute and buffer; photons carry — and the boundary between them lives where the crossover chart says, not where the architecture diagram is prettiest. The final lesson tours the places where that boundary has settled — and where it is moving.