MigSpike: A Migration Based Algorithms and Architecture for Scalable Robust Neuromorphic Systems

While conventional hardware neuromorphic systems usually consist of multiple clusters of neurons that communicate via an interconnect infrastructure, scaling up them confronts the reliability issue when faults in the neuron circuits and synaptic weight memories can cause faulty outputs. This work presents a method named
MigSpike
that allows placing spare neurons for repairing with the support of enhanced migrating methods and the built-in hardware architecture for migrating neurons between nodes (clusters of neurons).
MigSpike
architecture supports migrating the unmapped neurons from their nodes to suitable ones within the system by creating chains of migrations. Furthermore, a max-flow min-cut adaptation and a genetic algorithm approach are presented to solve the aforementioned problem. The evaluation results show that the proposed methods support recovery up to 100% of spare neurons. While the max-flow min-cut adaption can execute milliseconds, the genetic algorithm can help reduce the migration cost with a graceful degradation on communication cost. With a system of 256 neurons per node and a 20% fault rate, our approach minimizes the migration cost from remapping by 10.19× and 96.13× under Networks-on-Chip of
$4\times 4$
(smallest) and
$16\times 16 \times 16$
(largest), respectively. The Mean-Time-to-Failure evaluation also shows an approximate 10× of lifetime expectancy by having a 20% spare rate.