SpikON: A Dual-Parallel Efficient Accelerator for Online Spiking Neural Networks

SpikON is the first algorithm-hardware co-design framework for online supervised learning of Spiking Neural Networks (SNNs). Through techniques like learnable temporal thresholds and cascaded computation reuse, it achieves a 32.2% reduction in training latency and a 35% decrease in energy consumption, while enabling order-of-magnitude improvements in throughput and energy efficiency on edge devices. This project has been accepted by ISLPED 2026, and the code is open-sourced on GitHub (https://github.com/peilin-chen/SpikON).

As the third generation of neural networks, Spiking Neural Networks (SNNs) have great potential in energy efficiency due to their event-driven sparse computation characteristics. However, online learning faces two major bottlenecks:

1. Unsupervised online learning (e.g., STDP) is hardware-friendly but has low accuracy and poor scalability;
2. Supervised online learning (e.g., BPTT) has high accuracy but large memory overhead, and its computation mode does not match the parallel architecture of edge hardware, hindering deployment on resource-constrained devices.

The algorithm innovations of SpikON include:

• Learnable Temporal Threshold: Dynamically adjust threshold parameters to improve training stability and convergence speed;
• Temporally Scaled Weight Centering: Alleviate the gradient vanishing problem in deep SNNs and support deeper architectures;
• Cascaded Computation Reuse: Parallelize forward propagation and backpropagation across time steps (forward at step t while backpropagating at step t-1), reducing training latency and memory bandwidth requirements.

The core of the hardware architecture consists of Dual-Parallel Engine and SIMD-based SNN Core:

• The dual-parallel engine supports both forward and backpropagation simultaneously, maximizing resource utilization and reducing training latency by 32.2%;
• The SIMD core is optimized for the sparse characteristics of spikes, efficiently processing non-zero spike events, and uses a custom instruction set to accelerate specific operations.

Experimental results verify the performance advantages:

• Algorithm level: Under the premise of maintaining accuracy, training latency is reduced by 32.2% and energy consumption by 35%;
• System level: Compared to Apple M4 GPU, throughput is increased by 7.2x and energy efficiency by 11.5x; compared to TPU-like accelerators, throughput is increased by 26.8x and energy efficiency by 15.8x, making online SNN learning practical on edge devices.

The technical foundation relies on the PyTorch framework and the SpikingJelly open-source library, and references the efficient backpropagation algorithm from SLTT (ICCV2023). The open-source ecosystem facilitates researchers' reproduction and verification, as well as engineers' secondary development, promoting iterative improvements in the SNN community.

The application prospects are broad:

• Edge Intelligence: Real-time online learning for smart homes, wearable devices, etc., protecting privacy and reducing latency;
• Adaptive Systems: Continuous adaptation to new environments for robot navigation, autonomous driving perception, etc.;
• Low-Power Training: Providing new ideas for green AI. SpikON is a milestone in the transition of SNN online learning from theory to engineering practice, and is expected to promote the development of SNN accelerators.