A deterministic neuromorphic architecture with scalable time synchronization

Area of Science:

• Neuromorphic Engineering
• Computational Neuroscience
• Artificial Intelligence

Background:

• Custom integrated circuits model biological neural networks for brain computation research.
• Time synchronization is crucial for reproducibility and hardware-software equivalence in these systems.
• Existing global synchronization protocols hinder scalability.

Purpose of the Study:

• To develop a decentralized and scalable neuromorphic architecture named NeuroScale.
• To enable efficient large-scale network simulations without global coordination.
• To explore new artificial neural network architectures and learning rules.

Main Methods:

• NeuroScale employs local, aperiodic synchronization for determinism.
• Cores integrate compute and memory for neural and synaptic processes.
• Spike-based communication across a routing mesh with distributed event-driven synchronization.

Main Results:

• NeuroScale demonstrates scalability advantages over global barrier synchronization methods.
• The architecture supports modeling of spike filtering, subthreshold dynamics, and online Hebbian learning.
• Comparison with IBM TrueNorth and Intel Loihi highlights NeuroScale's benefits for large systems.

Conclusions:

• NeuroScale offers a scalable solution for neuromorphic computing by utilizing decentralized synchronization.
• This architecture facilitates the study of complex brain computations and advanced AI.
• The findings pave the way for more efficient and larger-scale neuromorphic systems.