Optical interconnects in AI clusters [/ˈɑptɪkəl ˈinterkonnekts ɪn ˈeɪaɪ ˈkləstərz/] n - A large AI training cluster is, to the networking eye, one enormous fabric that must let tens of thousands of accelerators exchange gradients almost as if they shared memory. The links that bear that traffic are increasingly optical, built on Silicon photonics and moving toward Co-packaged optics.
Why Light, and Why Now
Three pressures arrived at once:
- Bandwidth per accelerator [/ˈbandwidth ˈper ˈakselerator/] n - keeps climbing faster than copper can carry it at a reasonable reach.
- Reach. [/ˈreakh/] n - Copper at these speeds is good for roughly a meter. The moment two GPUs that must talk sit in different racks, you are in optical territory.
- Power. [/ˈpower/] n - The energy spent on the interconnect is now large enough to appear in the facility’s Power usage effectiveness, so cutting it is a first-class design goal, not an afterthought.
The Optical Circuit Switch
A newer notion borrowed from hyperscale practice is the optical circuit switch (OCS) [/ˈoptikal ˈsirkuit ˈswitkh ˈ(oks)/] n - a switch that physically steers beams of light (often with tiny movable mirrors) rather than converting to electrical signals and back. Because it never touches the bits, an OCS is largely bit-rate agnostic and burns very little power. It lets operators reconfigure the fabric topology to match a training job’s communication pattern, and to route round a failed rack without re-cabling.
What the Fabric Asks of the Building
Dense optical fabrics concentrate enormous power into small volumes and demand very stable temperatures for the photonics. That is a direct push toward Direct-to-chip liquid cooling and toward higher-voltage in-rack power like 800-volt DC distribution. Network, power, and cooling can no longer be designed by separate teams; see AI factory power and cooling.
Part of The New Buildout.