Deadlocks in NoC-based Neuromorphic Systems

Abstract

Neuromorphic systems promise to enable sustainable AI by substantially reducing power, area, and latency. These systems typically comprise multicores hosting neurons and a NoC to communicate spikes, i.e., small messages about neuron activations. System-level deadlocks from stuck in-flight spikes can disrupt neuromorphic systems. This paper solves this issue by a novel model for neuromorphic message passing that yields mapping constraints and methods. In a case study, we validate our approach and demonstrate reasonable resource utilization.

Authors 4

1. Jan Moritz Joseph Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University,Aachen,Germany

2. Jörn Nöller
   Affiliation as printed

   RooflineAI GmbH,Aachen,Germany

3. José Cubero-Cascante
   Affiliation as printed

   RooflineAI GmbH,Aachen,Germany

4. Rebecca Pelke
   Affiliation as printed

   RooflineAI GmbH,Aachen,Germany