New open-source framework optimizes mixed-signal spiking neural networks for edge AI

Energy-efficient neuromorphic computing at the edge demands simulation tools that capture the non-ideal behavior of mixed-signal SNN hardware. Researchers from multiple institutions, led by Sayma Nowshin Chowdhury, have introduced an open-source framework that enables design space exploration for such systems. The framework supports multiple neuron models—Leaky Integrate-and-Fire (LIF), Hodgkin-Huxley (HH), and Axon-Hillock (AH)—alongside non-volatile analog synapses built from floating-gate transistors and ReRAM devices. Crucially, it incorporates device-level nonlinearities directly into PyTorch-based training and inference, allowing optimization of physical synaptic parameters rather than abstract weights. This approach closes the gap between simulation and real hardware, making the framework a practical tool for chip designers.

The framework was evaluated on standard neuromorphic benchmarks: N-MNIST, DVS Gesture, and Spiking Heidelberg Digits (SHD). For each model-dataset combination, it reports classification accuracy together with hardware-oriented metrics such as silicon area, power consumption, and quantization sensitivity. This dual reporting enables designers to compare trade-offs and identify neuron-synapse configurations that best satisfy application-specific constraints on accuracy, energy efficiency, area, and hardware fidelity. By enabling cross-layer design space exploration, the framework accelerates the development of mixed-signal SNN accelerators for edge AI applications, from wearable devices to autonomous sensors.