NeuroWave-GO: A Reinforcement Learning Framework for Scalable and Perception-Aware SNN-to-NoC Mapping

Large-scale spiking neural network (SNN)-to-network-on-chip (NoC) mapping faces two bottlenecks: topology-only graphs underrepresent temporal communication dynamics, causing a perception bottleneck, while neuron-level optimization scales poorly, causing a scalability bottleneck. We propose NeuroWave-GraphOrchestrator (GO), which addresses them through a cluster-level mapping hierarchy. NeuroWave groups neurons into cohesive functional clusters via physics-inspired wave propagation, reducing search complexity and tackling the scalability bottleneck. GO then places these clusters onto NoC cores with decoupled attention, combining a lightweight static query, dense dynamic context, and physical inductive biases to overcome the perception bottleneck. Across benchmarks from multilayer perceptrons (MLPs) to million-neuron SNNs, NeuroWave-GO improves communication quality and scales beyond the reach of conventional mappers. On ResNet-18, it reduces average packet latency by 23.8% and communication cost by 39.9% over competitive baselines. Its cluster-level design also makes optimization more efficient: NeuroWave runs up to $2.4\times $ faster than Metis, and the optimized RL pipeline increases training throughput by $2.34\times $ .