Bandwidth and Crosstalk Optimisation for Wavelength-Routed Optical Networks-on-Chip Using Pareto Simulated Annealing

Optical Networks-on-Chip (ONoCs) are gaining traction in integrated circuits due to increasing network demands. Wavelength-Routed ONoCs (WRONoCs), a specialised subtype, further enhance performance by reducing network latency. To achieve efficient communication, WRONoC designs aim to maximise bandwidth and minimise crosstalk – two competing objectives requiring careful balance. This work introduces an improved Pareto Simulated Annealing algorithm broadly applicable to multi-objective combinatorial problems. Key advancements include enhanced parallelism and using an adaptive temperature schedule to ensure consistency across all cost dimensions. We apply this algorithm to the optimisation of bandwidth and crosstalk in any WRONoC design, where it handles all objectives separately or together, with the combined optimisation yielding Pareto fronts for trade-off analysis. We also use a substitute for the $\max$ objective function that accelerates solution improvement and achieves better final results. Results demonstrate that our algorithm significantly outperforms state of the art methods in solution quality ($1.14\times$ higher bandwidth on average) and runtime ($11176\times$ lower on average) while scaling effectively to substantially larger problem sizes (up to $8\times$ more network nodes).