When are Spiking Neural Networks Sustainable? A Hardware-Aware Carbon Analysis of Convolutional and Spiking Neural Models

Spiking neural networks (SNNs) are widely promoted as energy-efficient alternatives to conventional neural networks, yet their sustainability on general-purpose hardware remains unclear. This paper presents a hardware-aware carbon analysis of convolutional neural networks (CNNs), ANN–SNN conversion models, and fully spiking neural networks trained using surrogate gradient backpropagation through time. Using the MNIST dataset, we measure training and inference time, accuracy, and CO2 emissions with CodeCarbon, complemented by spike-count-based theoretical energy estimation. Our results show that CNNs achieve the highest accuracy with negligible inference-time carbon cost on CPUs, whereas ANN–SNN models incur substantial inference overhead due to multi-timestep spike simulation, and fully spiking networks exhibit the highest training-phase emissions from temporal unrolling and surrogate optimization. Despite low theoretical spike energy, system-level measurements reveal a clear gap between algorithmic efficiency and carbon impact under software-based execution. These findings demonstrate that SNN sustainability is strongly hardware-dependent and becomes meaningful primarily on event-driven or neuromorphic platforms.