A photonic-accelerator perspective identifies the surrounding bottlenecks

A September Nature Photonics perspective maps where optical accelerators could fit into AI infrastructure. It separates general cloud workloads from specialized edge applications with different requirements for scale, latency and power.

For large systems, optical losses and conversion between electrical and optical signals can dominate the budget. At the edge, applications may benefit from rapid parallel processing but remain constrained by nonlinear operations, reconfiguration and physical size.

The analysis favors evaluating photonics inside a complete digital system. A fast optical multiplication stage is useful only if input preparation, memory access and readout preserve enough of that advantage for the actual workload.