A Resilient ReRAM crossbar-based PIM Design for SNN in Energy Harvesting Scenarios

For edge IoT devices, ambient energy harvesting offers a batteryless energy solution by collecting environmental energy. When using a ReRAM crossbar to simulate spiking neural networks (SNNs), a single ReRAM crossbar is often not capable of simulating large SNNs due to size limitations. Moreover, the fluctuating nature of ambient energy may disrupt SNN operations, leading to intermittent computing. To address these issues, this paper designs a multi-ReRAM crossbar network to simulate the SNN. On this basis, we design multiple activation solvers corresponding to different energy levels. Therefore, the operating status can be adjusted to a novel "Idle-Partially Activated-Full Activated" mode to maximize the conversion of environmental energy into execution scheduling. Furthermore, we propose a boundary task rescue strategy based on energy prediction, which adjusts the activation scale of the set of crossbars according to the result of energy prediction, so as to preserve incomplete tasks at the boundary of the power cycle and smoothly transition them to the next power cycle. The experimental results show that compared with the traditional operation mode of pulsed neural networks, this strategy greatly improves the throughput rate and energy efficiency of the system, and can effectively preserve the progress of boundary tasks.