A High-Energy-Efficiency Lightweight BNN Accelerator
A High-Energy-Efficiency Lightweight BNN Accelerator for Arrhythmia Detection
Abstract:
Arrhythmia poses a significant threat to human health. Accurate and real-time arrhythmia detection using elec-trocardiogram (ECG) signals is critical for medical diagnosis. With the rapid development of deep learning technologies, Deep Neural Networks (DNNs) have considerably improved the accuracy of arrhythmia monitoring. However, deploying these techniques on power-constrained wearable devices for real-time monitoring remains challenging. Lightweight designs have thus become essential for achieving low-power and high-throughput real-time detection with limited hardware resources. Binary Neural Networks (BNNs), as a promising lightweight method, exhibit significant potential in reducing hardware resource usage and enhancing inference speed. Nevertheless, existing binarized ECG detection accelerator designs have not yet conducted archi-tectural Design Space Exploration (DSE), and therefore do not sufficiently balance the trade-off between throughput and power consumption. To address this, we propose a highly energy-efficient one-dimensional BNN accelerator for ECG detection. First, we develop a hardware-friendly BNN algorithm by eliminating fully connected layers and nonlinear computations. Subsequently, for the first time, this work utilizes Timeloop to analyze the optimal mapping of the BNN algorithm onto hardware, identifies the optimal accelerator dataflow, and conducts in-depth optimization of both the dataflow and the accelerator hardware circuits based on the Field-Programmable Gate Array (FPGA) architecture. Finally, we introduce a prediction mechanism for fine-grained power control by dynamically disabling redundant computational logic. Experimental results demonstrate that the proposed accel-erator achieves a latency of 105.79 µs, power consumption of 160 mW, throughput of 112.2 GOPS, and energy efficiency of 701.51 GOPS/W at 100 MHz. Compared with state-of-the-art Xilinx FPGA-based ECG detection accelerators, our design improves energy efficiency by 15.
” Thanks for Visit this project Pages – Register This Project and Buy soon with Novelty “
A High-Energy-Efficiency Lightweight BNN Accelerator for Arrhythmia Detection
