Neuromorphic systems are scaling to thousands of near-memory cores, but aging can cause permanent faults in the neuron-update datapaths. A new arXiv preprint argues that these faults are a major obstacle to long-term deployment and proposes a recovery strategy that migrates computation rather than data.
According to the abstract, existing online recovery methods typically retire the faulty element. The new approach instead moves the computation that would have used the damaged datapath, leaving data in place. This distinction matters because moving data in a many-core system is expensive and can disrupt ongoing neural processing.
The preprint is listed as arXiv:2610.26985. Because only the abstract is available, the specific mechanism and overhead are not yet described, and no independent verification is included.