基于SOT-MRAM的超低功耗心电智能处理器关键技术研究

可穿戴心电监测设备被用于实时监测心律不齐等心脏疾病，而心电智能处理器是该设备的核心，具有超低功耗的要求，以保证设备维持长时间稳定的分类检测。但是，传统智能计算架构中数据不断在存储器和计算单元间移动，带来较大的动态功耗开销。同时，随着半导体工艺的进步，存储器的静态功耗比例不断升高，导致依托于传统技术的心电智能处理器难以满足可穿戴心电监测设备的超低功耗要求。本项目旨在解决心电智能处理器中计算和存储架构所面临的功耗瓶颈，利用新兴的自旋轨道矩存储器（SOT-MRAM）技术设计心电智能处理器的计算和缓存架构。但是，现有的SOT-MRAM技术针对通用计算和存储架构设计，没有根据心电信号特点进行设计，有冗余的功耗开销。并且心电信号分类算法模型在超低功耗限制下也将进一步优化，以满足心电信号监测的准确率要求。该项目对推动我国可穿戴心电设备智能化和国产化，具有重要的科学意义和前景。

Wearable Electrocardiograph (ECG) monitoring devices are employed to detect heart diseases such as arrhythmia in real time. In the ECG monitoring, ECG intelligent processor is the core element with low power requirement to prolong the stable monitoring for classification and detection. However, there are large data movements between computing unit and working memory in traditional intelligent computing architecture, inducing large dynamic power consumption. In addition, as the CMOS technology scaling down, the proportion of static power consumption of memory is increasing. Therefore, the ECG intelligent processor with traditional CMOS technology is difficult to meet the ultra-low power requirement of wearable ECG monitoring devices. In this proposal, we develop the ECG intelligent processor using Spin-Orbit Torque Magnetoresistance Random Access Memory (SOT-MRAM) to solve power issues in both computing and memory architectures. Unfortunately, current solutions with SOT-MRAM techniques are developed for universal computing and memory architectures. The ECG intelligent processor with current SOT-MRAM solutions brings redundancy power consumption due to without considering characteristics of ECG. Moreover, we provide optimization ECG classification algorithm to meet both ultra-low power and high accuracy requirements of ECG intelligent processor. Our proposal has important scientific significance and prospect for promoting the intellectualization and localization of China's wearable ECG devices.