Hardware/Software Co-Design of Ultra-Low Power Biomedical Monitors

Ongoing changes in world demographics and the prevalence of unhealthy lifestyles are imposing a paradigm shift in healthcare delivery. Nowadays, chronic ailments such as cardiovascular diseases, hypertension and diabetes, represent the most common causes of death according to the World Health Organization. It is estimated that 63% of deaths worldwide are directly or indirectly related to these non-communicable diseases (NCDs), and by 2030 it is predicted that the health delivery cost will reach an amount comparable to 75% of the current GDP. In this context, technologies based on Wireless Sensor Nodes (WSNs) effectively alleviate this burden enabling the conception of wearable biomedical monitors composed of one or several devices connected through a Wireless Body Sensor Network (WBSN). Energy efficiency is of paramount importance for these devices, which must operate for prolonged periods of time with a single battery charge. In this thesis I propose a set of hardware/software co-design techniques to drastically increase the energy efficiency of bio-medical monitors. To this end, I jointly explore different alternatives to reduce the required computational effort at the software level while optimizing the power consumption of the processing hardware by employing ultra-low power multi-core architectures that exploit DSP application characteristics. First, at the sensor level, I study the utilization of a heartbeat classifier to perform selective advanced DSP on state-of-the-art ECG bio-medical monitors. To this end, I developed a framework to design and train real-time, lightweight heartbeat neuro-fuzzy classifiers, detail- ing the required optimizations to efficiently execute them on a resource-constrained platform. Then, at the network level I propose a more complex transmission-aware WBSN for activity monitoring that provides different tradeoffs between classification accuracy and transmission volume. In this work, I study the combination of a minimal set of WSNs with a smartphone, and propose two classification schemes that trade accuracy for transmission volume. The proposed method can achieve accuracies ranging from 88% to 97% and can save up to 86% of wireless transmissions, outperforming the state-of-the-art alternatives. Second, I propose a synchronization-based low-power multi-core architecture for bio-signal processing. I introduce a hardware/software synchronization mechanism that allows to achieve high energy efficiency while parallelizing the execution of multi-channel DSP applications. Then, I generalize the methodology to support bio-signal processing applications with an arbitrarily high degree of parallelism. Due to the benefits of SIMD execution and software pipelining, the architecture can reduce its power consumption by up 38% when compared to an equivalent low-power single-core alternative. Finally, I focused on the optimization of the multi-core memory subsystem, which is the major contributor to the overall system power consumption. First I considered a hybrid memory subsystem featuring a small reliable partition that can operate at ultra-low voltage enabling low-power buffering of data and obtaining up to 50% energy savings. Second, I explore a two-level memory hierarchy based on non-volatile memories (NVM) that allows for aggressive fine-grained power gating enabled by emerging low-power NVM technologies and monolithic 3D integration. Experimental results show that, by adopting this memory hierarchy, power consumption can be reduced by 5.42x in the DSP stage.