Energy and Latency

Passive propagation through fabricated optical elements is described as requiring no electronic computational energy in the optical path. DriftXpress does not materially change inference latency because it uses the same one-step generator after training. The paper distinguishes computational energy during passive propagation from total end-to-end system power. The comparison with electronic systems is architectural rather than task-normalized. Per-sample latency was approximately 0.177 ms across the evaluated datasets. End-to-end latency and energy are not fully measured because optical propagation time excludes source, detection, I/O, and electronic processing overhead. After training, inference remains a single UNet forward pass. Total system energy is not zero because sources, photodetectors, and control electronics add overhead.