A configurable streaming spiking neural network accelerator with decoupled pixel-level and output-channel parallelism...

Streaming spiking neural network (SNN) accelerators are widely adopted on edge platforms for their deterministic, low-latency inference. Realizing their full efficiency, however, requires three properties to hold simultaneously: a router-free streaming dataflow, joint exploitation of temporal and spatial sparsity within that dataflow, and configurable parallelism that can adapt to heterogeneous layers and hardware budgets. Existing streaming SNN accelerators typically achieve at most two of these properties: pixel-level and output-channel parallelism are bound to a single fixed operating point, limiting efficient mapping across heterogeneous layers in automatic modulation classification (AMC) workloads. This paper presents a configurable streaming SNN accelerator that satisfies all three properties by explicitly decoupling pixel-level multiple matrix-vector multiplication (MMV) parallelism from output-channel parallelism. The decoupling is realized within a weight-priority gated one-to-all product (GOAP) dataflow by deterministic offline scheduling, preserving router-free streaming execution while exploiting temporal and spatial sparsity. The proposed architecture is implemented on a Xilinx Virtex-7 field-programmable gate array (FPGA) and evaluated using the RadioML 2016 dataset and a compatible subset of RadioML 2018 under multiple sparsity and quantization settings. Experimental results show that, under comparable end-to-end latency, configurable parallelism enables effective layer-wise latency balancing and substantial hardware-resource savings while sustaining high throughput and classification accuracy. More broadly, the same accelerator description can be retargeted across operating points spanning more than an order of magnitude in hardware cost, providing a key enabler for deploying streaming SNNs at scale across diverse edge hardware platforms.