SNN vs. CNN Implementations on FPGAs: An Empirical Evaluation

Convolutional Neural Networks (CNNs) are widely employed to solve various problems, e.g., image classification. Due to their compute- and data-intensive nature, CNN accelerators have been developed as ASICs or on FPGAs. The increasing complexity of applications has caused resource costs and energy requirements of these accelerators to grow. Spiking Neural Networks (SNNs) are an emerging alternative to CNN implementations, promising higher resource and energy efficiency. The main research question addressed in this paper is whether SNN accelerators truly meet these expectations of reduced energy demands compared to their CNN equivalents when implemented on modern FPGAs. For this purpose, we analyze multiple SNN hardware accelerators for FPGAs regarding performance and energy efficiency. We also present a novel encoding scheme of spike event queues and a novel memory organization technique to improve SNN energy efficiency further. Both techniques have been integrated into a state-of-the-art SNN architecture and evaluated for MNIST, SVHN, and CIFAR-10 data sets and corresponding network architectures on two differently sized modern FPGA platforms. A result of our empirical analysis is that for complex benchmarks such as SVHN and CIFAR-10, SNNs do live up to their expectations.