E-DOSA: Efficient Dataflow for Optimising SNN Acceleration
Hemangee K. Kapoor, Imlijungla Longchar, Binayak Behera · 2025
Machine Learning has seen significant growth re-cently, from revolutionising industries in healthcare to finance. As the field of neuromorphic engineering keeps progressing, Spiking Neural Networks (SNNs) have also gained significant attention for their potential to create brain-like computational systems, offering prospects for more advanced, efficient, and adaptive artificial intelligence models. SNNs, inspired by real neurons, represent a distinct type of artificial neural network (ANN) that communicates through discrete spikes. This spiking mechanism allows for time-based information processing and can lead to a more energy-efficient neural computation. Compared to traditional ANNs, SNNs have more computation and add to sparsity due to the addition of a time dimension. Having an effective inference architecture and dataflow is im-perative to perform the tasks efficiently and, at the same time, being energy efficient. In this paper, we propose a dataflow that splits the inputs across time-batches and divides the processing of kernels across different PEs. This helps to reduce memory access and enables data reuse. Using a 2D systolic array of PEs, we achieve an average improvement of 1.2x and 6.6x over existing designs.