Abstract
Neuromorphic processors leverage brain-inspired spike-based sparse computation to gain energy efficiency for edge-node intelligence. Yet, achieving high-accuracy on-chip learning for in-situ self-adaptation at a low cost and a low latency poses a key challenge. To address this, we present the 4th generation of our MorphBungee processors. The processor adopts our integer e-prop algorithm for high-accuracy, low-cost on-chip SNN training. It employs a vector-scalar dual-core architecture with 2 levels of pipelines and the spike/error-event-driven paradigm to ensure a high sample throughput. A logically transposable memory enables the vector core’s flexible parallelism along the neuron- or the synapse-dimension in different processing phases. We prototyped our design on a very-low-cost Xilinx Zynq-7010 FPGA chip, and demonstrated its overall performance superiority over other related works, in terms of on-chip learning accuracy, processing speed, resource cost and energy efficiency.
Funding Information
  • Major Research Plan of the National Natural Science Foundation of China (92464103)
  • Key Program of the National Natural Science Foundation of China (62334008)