Silicon photonics [/ˈsɪləkən ˈfotoniks/] n - Silicon photonics builds optical components (waveguides, modulators, photodetectors) from the same silicon used for logic, so that light may be generated, steered, and received on a chip rather than only along long fiber runs. For years it dwelt at the edges of the data center, in the pluggable optical transceivers that stop fiber at a switch. It appears now in pages about power for one reason: the energy cost of moving a bit has stopped falling as fast as the energy cost of computing one.

Why These Pages Concern Themselves With It

A modern AI training fabric moves staggering quantities of data between accelerators. With traditional electrical links, a growing share of a rack’s energy budget goes not to computation but to pushing bits across copper and through electrical-to-optical conversions at the switch faceplate. That energy becomes heat, and heat must be removed; an interconnect problem quietly becomes a Power usage effectiveness problem and a cooling problem.

Silicon photonics attacks this by shortening the electrical path: rather than driving a signal several inches across a board to a faceplate transceiver, the optics are drawn much closer to the chip that produced the data. The next step along that road is Co-packaged optics.

Of the Pieces

• Waveguides [/ˈwævɛgʌɪdɛs/] n route light across the die with very low loss.
• Modulators [/ˈmɒdʌlætɒrs/] n imprint data onto a beam by altering the silicon’s refractive index electrically.
• Photodetectors [/ˈphɒtɒdɛtɛktɒrs/] n (usually germanium on silicon) turn the received light back into current.
• Light source. [/ˈlɪght ˈsæʊrkɛ/] n Silicon is a poor light emitter, so the laser is typically a separate indium-phosphide chip bonded on or coupled in. That external laser remains one of the field’s persistent headaches of reliability and power.

Wherein the Difficulty

Photonic devices are temperature-sensitive; a modulator’s tuning drifts with a few degrees of change, so they require either active thermal tuning (which costs power) or athermal designs. That sensitivity is one more reason the industry leans into Direct-to-chip liquid cooling, which holds junction temperatures far more stable than air ever could.

Part of The New Buildout. See also Optical interconnects in AI clusters.