SNN placement tools have been solving the wrong problem. Graph partitioning and pairwise placement costs are the standard formulation, but a spike from a source neuron delivers to a set of postsynaptic destinations (routes to multiple destinations can share mesh links). Optimizing edge cuts penalizes what should be shared. M-HySMap starts from the correct abstraction: directed hyperedges weighted by profiled activity, with optimization over actual route reuse on the mesh.
The implementation uses activity-aware graph seeds, exact incremental gain evaluation via locality (moving one neuron only changes its own source-rooted hyperedge and its predecessors), and a portfolio of placement neighborhoods that preserves the best incumbent. On a 115-job evidence suite across Potjans-inspired recurrent SNNs and mesh NoCs from 4x4 to 7x7, M-HySMap reduces routed multicast hops 10.6-19.6% versus Activity+QAP and 19.7-41.1% versus Edge+QAP. Incremental updates run 4.7-12.7x faster than full recomputation while matching it to numerical precision.
The specific teams this unblocks are the ones building or deploying neuromorphic inference hardware for sparse spiking workloads at the edge: automotive sensor fusion, DVS-based vision, and the growing class of GNN workloads that map cleanly to spiking execution. The broader point is that placement tooling for neuromorphic systems has been adapting graph-neural-network-era tools without asking whether the underlying metric translates. It does not. M-HySMap is the first explicit fix; others in the neuromorphic EDA stack likely have the same problem waiting to be named.