Unifying Feed-Forward and Weight-Gradient: An Energy-Efficient Architecture for On-Chip SNN Learning
Abstract
Spiking neural networks (SNNs) enable efficient on-chip learning but face a critical hardware challenge due to the computational asymmetry between small-kernel convolution operations in the feed-forward (FF) stage and large-kernel ones in the weight-gradient (WG) stage. Directly mapping these conflicting patterns onto a static architecture disrupts data locality, typically necessitating either disjoint processing units or redundant data movement, which severely compromises energy efficiency. This brief presents an architecture for unified learning in SNN (AULiS), a sparsity-aware accelerator supporting low-cost BFloat16 local learning (LL) via reconfigurable engines. Its core innovation is an algorithm-architecture codesign. We reformulate the WG stage into broadcast multiplications and map it to an output-stationary (OS) dataflow, newly applied to WG, complementing the weight-stationary (WS) dataflow used for FF. This strategy matches the distinct data-reuse characteristics of each stage: locking weights during FF and weight gradients during WG, allowing a single reconfigurable engine to process both, thereby minimizing data access and maximizing compute density and energy efficiency. Implemented in a 28-nm process, AULiS occupies 2.24 mm<sup>2</sup> and achieves this unification with only a 3.4% area overhead. The architecture reaches 23.0 TFLOPS/W in inference and 22.3 TFLOPS/W in training, in dense-equivalent FLOPs at 95% zero sparsity. In a fair comparison, this represents a <inline-formula> <tex-math notation="LaTeX">$3.78\times $ </tex-math></inline-formula> training and <inline-formula> <tex-math notation="LaTeX">$1.50\times $ </tex-math></inline-formula> inference improvement in normalized energy efficiency over prior SOTA while delivering up to <inline-formula> <tex-math notation="LaTeX">$10.64\times $ </tex-math></inline-formula> higher normalized compute density. The high and balanced efficiencies prove that learning can be unified with high-performance inference without efficiency compromise for adaptive edge artificial intelligence (AI).
