| Abstract: |
The rapid expansion of edge computing applications necessitates energy-efficient AI solutions which can work under strict power constraints while embracing high inference accuracy. In this paper, we introduce a holistic hardware-software co-design for spiking neural networks (SNNs) on memristive crossbar arrays with the highest energy efficiency ever reported in literature towards edge intelligence applications. Our framework is built on three key innovations: (1) a novel Leaky Integrate-and-Fire with Adaptive Threshold (LIF-AT) neuron model which reduces spike activity by 43.7% through dynamically selecting its threshold based on network-level statistics, leading to significant energy savings without compromising classification accuracy; (2) a hardware-aware training approach that accounts for memristor device non-idealities such as conductance drift, stuck-at faults, and programming variability which achieves within 1.8% of software ideal accuracy even in the presence of 5% device fault rates; and (3) an optimized weight mapping strategy using ternary weight quantization accompanied by asymmetric thresholds reducing analog-to-digital converter (ADC) resolution requirements from 8 bits down to 4 bits while achieving the state-of-the-art network representation accuracy of $97.3\%$ CIFAR-10 classification. We construct and evaluate 1T1R memristive crossbar arrays with the HfO 2 -based resistive switching devices, achieving a 32×32 array fabrication with a successful yield of 97.2% and programmable conductance levels up to 4 bits. System-level experiments on image classification (CIFAR-10, ImageNet-subset), speech recognition (Google Speech Commands) and hand gesture recognition (DVS-Gesture) benchmarks show that our framework achieves 127.3 TOPS/W energy efficiency, which is a 23.4× improvement over state-of-the-art GPU implementations and 8.7× better performance than recent neuromorphic accelerators. The complete framework, including training algorithms, hardware mo |
