| |
Congyang Li, Nabil Imam, and Rajit Manohar
Custom integrated circuits modeling biological neural networks serve
as tools for studying brain computation and platforms for exploring
new architectures and learning rules of artificial neural networks.
Time synchronization across network units is an important aspect of
these designs to ensure reproducible results and maintain
hardware-software equivalence. Current approaches rely on global
synchronization protocols, which fundamentally limit system scalability.
To overcome this, we develop NeuroScale, a decentralized and scalable
neuromorphic architecture that uses local, aperiodic synchronization
to preserve determinism without global coordination. Cores of co-localized
compute and memory elements model neural and synaptic processes, including
spike filtering operations, subthreshold neural dynamics, and online Hebbian
learning rules. Multiple cores communicate via spikes across a routing mesh,
using distributed event-driven synchronization to efficiently scale to large
networks. We compare this synchronization protocol to the global barrier
synchronization approaches of IBM TrueNorth and Intel Loihi, demonstrating
NeuroScale's advantages for large system size.
|
|
