NeMo

Abstract

Neuromorphic computing is a broad category of non–von Neumann architectures that mimic biological nervous systems using hardware. Current research shows that this class of computing can execute data classification algorithms using only a tiny fraction of the power conventional CPUs require. This raises the larger research question: How might neuromorphic computing be used to improve application performance, power consumption, and overall system reliability of future supercomputers? To address this question, an open-source neuromorphic processor architecture simulator called NeMo is being developed. This effort will enable the design space exploration of potential heterogeneous compute systems that combine traditional CPUs, GPUs, and neuromorphic hardware. This article examines the design, implementation, and performance of NeMo . Demonstration of NeMo ’s efficient execution using 2,048 nodes of an IBM Blue Gene/Q system, modeling 8,388,608 neuromorphic processing cores is reported. The peak performance of NeMo is just over ten billion events-per-second when operating at this scale.

SHIP: a computational framework for simulating and validating novel technologies in hardware spiking neural networks

“…As ED algorithms are in principle strictly sequential, the use of parallel computing techniques is contingent on the synchronization between modeled components. This is a self-contained computational problem, finding no simple solution but being actively researched (see e.g., Plagge et al, 2018 ; Pimpini et al, 2022 ). In summary, event-driven approaches carry out calculations only where the events take place, but each event adds to the model operations the overhead of a set of corollary functions.…”