eMX: A configurable mixed-precision CNN accelerator for low-power wearable ECG classification

Electrocardiogram (ECG) classification is essential for wearable healthcare because on-device inference reduces communication overhead and clinicians’ workload. Compact one-dimensional convolutional neural networks (1-D CNNs) have improved ECG classification accuracy; however, existing CNN accelerators still suffer from large parameter memories because most support only standard convolution or fixed high-precision arithmetic. We present eMX, a configurable mixed-precision CNN accelerator for 1-D CNN-based ECG classification with an extremely small model-memory footprint. Three key ideas are developed. First, we develop a compact ECG model supporting standard, pointwise, depthwise, and multi-kernel depthwise convolutions, achieving an F1 score of 98.88% on the MIT-BIH dataset with only 4921 parameters. Second, eMX supports W8A8, W4A8, and mixed W4/8A8 execution on a unified architecture, requiring only 2.84 KB for the mixed W4/8A8 model. Third, eMX introduces a configurable hardware architecture driven by 32-bit instructions, supporting convolution types through programmable kernel size, stride, precision, feature-map routing, and residual/branch operations. Compared with existing FPGA-based ECG accelerators, eMX achieves an area efficiency of 147.15 GOP/s/MeLUT and improves ADP by up to 167.9 × over 1-D CNN designs. Furthermore, ASIC synthesis in 45 nm technology achieves 0.581 mm 2 area and up to 1.76 × lower normalized energy ( μ J/classification) than existing works.