Quantify advantages of photonic probabilistic computing engines over digital computers

Determine which specific benefits photonic probabilistic computing engines can achieve relative to digital computers, by rigorously benchmarking sampling speed, energy efficiency, accuracy, scalability, and cost across representative tasks (e.g., stochastic neural network inference, Boltzmann sampling for spin-glass and molecular simulations, and probabilistic optimization). Establish application domains and conditions under which photonic probabilistic computing provides clear, reproducible performance improvements over conventional digital approaches.

Background

Photonic probabilistic computing engines aim to accelerate sampling tasks using physical stochastic processes (e.g., optical chaotic oscillations, bistable devices, parametric down-conversion) with potential gains in speed and energy relative to pseudorandom number generation on digital systems.

Despite promising demonstrations, these systems are in early stages. The paper notes the need to identify real-world use cases and to establish the extent and nature of benefits over digital computing through careful benchmarks and performance analyses.