NEURAL: An Elastic Neuromorphic Architecture with Hybrid Data-Event Execution and On-the-fly Attention Dataflow

2025-09-18

AI Summary

Existing SNN hardware suffers from high latency and low energy efficiency due to spike sparsity and multi-timestep execution. This paper proposes NEURAL, a reconfigurable brain-inspired architecture based on a hybrid data- and event-driven execution paradigm: it decouples sparsity-aware processing from neuron computation to enable online execution of spike-based QKFormer; introduces elastic FIFO scheduling and a Window-to-First-Spike-Time (W2TTFS) conversion mechanism to achieve dynamic attention fusion without dedicated hardware; and leverages knowledge distillation to construct a high-accuracy, single-timestep, fully spiking SNN. FPGA evaluation demonstrates that NEURAL reduces resource consumption by 50% and improves energy efficiency by 1.97× over state-of-the-art SNN accelerators, while achieving +3.20% and +5.13% accuracy gains on CIFAR-10 and CIFAR-100, respectively, using VGG-11.

Technology Category

• Cognitive Modeling & Cognitive Systems: Neural Spike Coding
• Machine Learning: Hardware-aware ML
• Natural Language Processing: Learning & Optimization for NLP

Application Category

• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Economics, Online Markets and Human Computation: Architectures and workflows that use LLMs for crowd work

Abstract

Spiking neural networks (SNNs) have emerged as a promising alternative to artificial neural networks (ANNs), offering improved energy efficiency by leveraging sparse and event-driven computation. However, existing hardware implementations of SNNs still suffer from the inherent spike sparsity and multi-timestep execution, which significantly increase latency and reduce energy efficiency. This study presents NEURAL, a novel neuromorphic architecture based on a hybrid data-event execution paradigm by decoupling sparsity-aware processing from neuron computation and using elastic first-in-first-out (FIFO). NEURAL supports on-the-fly execution of spiking QKFormer by embedding its operations within the baseline computing flow without requiring dedicated hardware units. It also integrates a novel window-to-time-to-first-spike (W2TTFS) mechanism to replace average pooling and enable full-spike execution. Furthermore, we introduce a knowledge distillation (KD)-based training framework to construct single-timestep SNN models with competitive accuracy. NEURAL is implemented on a Xilinx Virtex-7 FPGA and evaluated using ResNet-11, QKFResNet-11, and VGG-11. Experimental results demonstrate that, at the algorithm level, the VGG-11 model trained with KD improves accuracy by 3.20% on CIFAR-10 and 5.13% on CIFAR-100. At the architecture level, compared to existing SNN accelerators, NEURAL achieves a 50% reduction in resource utilization and a 1.97x improvement in energy efficiency.

Problem

• Addressing spike sparsity and multi-timestep execution inefficiencies in SNN hardware
• Implementing hybrid data-event execution paradigm for neuromorphic computing
• Enabling on-the-fly attention dataflow without dedicated hardware units

Innovation

• Hybrid data-event execution paradigm
• On-the-fly attention dataflow execution
• Knowledge distillation training framework