Workshop: One Switch Does Not Fit All: Do Training and Inference Demand Different Optical Switching Paradigms?

Optical switching for AI has moved from research to hyperscale deployment; yet the field is fragmenting fast. MEMS, silicon photonics, and LCoS-based OCS compete across incompatible tradeoff spaces, while AI workloads are diverging: training tolerates millisecond reconfiguration; inference demands sub-microsecond responsiveness that OCS alone cannot serve.

Meanwhile, copper is rapidly giving way to optics across the intra- to inter-rack boundary, raising urgent questions about which switching paradigm and implementation technology gets deployed at scale.

This workshop stages a direct confrontation between competing approaches, using training vs. inference requirements as the stress test.