Abstract
The inherent high sparsity of the spiking neural networks (SNNs) and the low power consumption brought by the event-driven computing characteristics suit edge devices with extremely high energy efficiency requirements. On resource-constrained mobile devices, we also require memory saving. Unlike conventional artificial neural networks, SNNs are suitable for processing complex temporal data. However, computing in the time dimension requires repeated access to the data for multiple time steps, resulting in high energy consumption. We propose Temporally Parallel Weight-Friendly (TPWF) dataflow, which reduces energy consumption through parallel computing across time steps. At the same time, considering the high sparsity of the spike event, this paper proposes a sparse aware strategy, which can realize high-energy-efficiency membrane potential accumulation calculation by the neuron burst weight search circuit. Furthermore, this paper proposes an efficient synaptic memory structure to reduce hardware resource usage while maintaining performance and network size. Use run-length encoding to record weights, realize synaptic connections that can support different configurations, such as sparse connection networks, and save a lot of memory.Using a fully connected 256-128-128-10 network to classify 16×16 MNIST training images achieves energy per synaptic operation (SOP) of 0.2pJ, up to 1.9x speedup, and 2x reduction in memory accesses.