A Programmable Systolic-Array AI Accelerator System with High-Performance Model Quantization and Heart Disease Classification Algorithm Design

• Published in: IEEE International Symposium on Circuits and Systems proceedings pp. 1 - 5
• Main Authors: Wang, Kuan-Cheng, Ku, Ming-Yueh, Lee, Shuenn-Yuh, Chen, Ju-Yi
• Format: Conference Proceeding
• Language: English
• Published: IEEE 25.05.2025
• Subjects: Arrhythmia; Artificial Intelligence Accelerator; Classification algorithms; Computational modeling; Computer architecture; Convolution Neural Network; Convolutional neural networks; Diseases; Electrocardiography; Heart valve diseases; Phonocardiography; Programmable; Quantization (signal); Systolic array; Systolic arrays
• ISSN: 2158-1525
• DOI: 10.1109/ISCAS56072.2025.11043735

This work introduces a heart disease classification system. The system includes electrocardiography (ECG) arrhythmia classification and phonocardiography (PCG) heart-valve diseases classification algorithm, achieving 97.4% and 99.1% accuracy. Additionally, the paper presents a procedure for lightweight convolutional neural network (CNN) model quantization with an 8-bit fix-point and 0.1% accuracy loss. Furthermore, this study proposes a programmable artificial intelligence (AI) accelerator with an application-specific instruction set processor (ASIP) and systolic array architecture to achieve high-performance computing. Moreover, we introduce a matrix mapping unit (MMU) and the pipeline state register (PSR) to facilitate switching between CNN and matrix multiplication, resulting in a reduction of over 50% in timing overhead. The chip is implemented on Xilinx's PYNQ-Z2 and achieves a power consumption of 106 mW, with a classification latency of 6.8ms / 21ms (arrhythmia/valve diseases).