MorphBungee: An Edge Neuromorphic Chip for High-Accuracy On-Chip Learning of Multiple-Layer Spiking Neural Networks

Abstract

Spiking neural networks and neuromorphic systems have attracted ever increasing interests recently, due to their high computational efficiency by imitating the functional mechanism of cerebral cortex. However, endowing low-cost neuromorphic chips with real-time high-accuracy on-chip learning plasticity for edge applications is still challenging. In this work, we present a digital edge neuromorphic chip for real-time high-accuracy on-chip multi-layer SNN learning in visual recognition tasks. It employs a hierarchical multi-core architecture, a dynamically reconfigurable array parallelism and a quasi-event-driven scheme to improve processing speed. A prototype chip with a core area of 10.39 mm <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> was fabricated using a 65-nm 1P9M CMOS process, and typically achieved a real-time speed of 87 frames/s and a power dissipation of 106 mW at an 83 MHz clock rate when training a 4-layer fully-connected SNN. Our chip attained comparably high recognition accuracies of 96.29%, 84.95%, 86.13%, 85.07% and 100% on the MNIST, Fashion-MNIST, ETH-80, MNIST-DVS and Poker-DVS datasets, respectively, with an energy efficiency of 97 pJ/SOP for learning and 30 pJ/SOP for inference.