SBGIP: Sparse Binary Guided Inner Product for Energy‐Aware CNN Inference Acceleration on Microcontrollers

ABSTRACT

This paper proposes sparse binary guided inner product (SBGIP), a software‐level approximation method for convolutional neural network (CNN) inference on low‐power microcontrollers (MCUs) without SIMD or DSP extensions. SBGIP builds upon the approximate inner product idea but replaces AIP's 8‐bit contiguous subset with a dispersed, sparse 1‐bit sign subset. For each output neuron, a low‐cost Stage‐1 sign estimate decides whether the subsequent ReLU will activate; only positively gated neurons trigger Stage‐2, which recomputes the exact inner product using the original full‐precision weights. We formalise the method, derive a minimum‐subset‐size criterion ≳ 32 and clarify activation reuse, weight precision and per‐output scheduling. Evaluation on six models—LeNet‐5, AlexNet, GoogLeNet (no‐BN), ResNet‐18 (no‐BN), MobileNetV2 (no‐BN) and DS‐CNN‐S for keyword spotting—shows 16%–71% full‐model energy savings versus FP32 under a 45∼nm Horowitz model. Bare‐metal STM32F429IGT6 firmware achieves up to 5.53× lower latency than a vanilla TensorFlow Lite Micro int8 baseline, with board‐level power measurements confirming 7%–43% energy savings.