Process scaling has enabled the emergence of many small systems with diverse functions. Expecting a lot more IoT devices around us in the future, it is becoming more of a concern to lower the manufacturing and operating costs of each device, which include costs for design, production and installation, and also an end-user's overheads for maintaining IoT devices such as periodic charging and troubleshooting. I am particularly interested in developing an autonomous system of remote devices that is powerful, reliable and cheap enough so that an end-user can “install and forget” such devices to easily build and maintain his/her own IoT network.
There are still many challenges towards this research goal, and my main research objectives are to address these challenges and to apply the research outcomes to real IoT applications.
Machine Learning on Edge Devices
Data transfer between remote devices and a base station is costly in terms of both energy and latency, so it is necessary to preprocess collected data before wireless transmission. There are already many techniques such as simple filtering, feature extraction, and compressed sensing to preprocess and compress the row dara. However, compared to them, machine learning can process more various types of data by a single algorithm, which allows us to make a unified data processing accelerator that can be widely used in many device regardless of specific data type.
Energy-Efficient Sensors in Advanced Technologies
Sensing environmental variables is one of the most essential functions in many IoT applications, but sensor interfaces are more difficult to scale down in terms of area, cost, and power consumption compared to digital circuits. Thermal noise, process variations, and the reduction of output swing and intrinsic gain in advanced processes are the main factors that restrain the performance of analog circuits and render mere scaling unfavorable.
One of the approaches to mitigate these issues is to modify the topology of conventional analog circuits to look more like digital circuits, so that they fully benefit from process scaling. This scheme can be applied to many different types of sensors and ADCs, and can be easily combined with other digital-oriented techniques such as dynamic tuning of sensor's operating rate and sensitivity. In addition, its simpler and more robust structure brings down the cost for design.
Autonomous Power Management for Self-Powered Devices
Because of the limited energy and power availability in a small and remote IoT device, improving efficiency in energy harvesting and power management is essential to extend overall system operating time. In addition, as dynamic voltage and frequency scaling (DVFS) is becoming more popular and fine-grained on both low-power and high-performance applications for higher power efficiency, the number of required power regulators as well as their design and implementation costs are increasing.
Inductive switching power conversion is widely used for various power conversion applications, but the need for big off-chip inductors renders them less favorable in small, low-power systems. Conversely, switched-capacitor (SC) DC-DC converters can be fully integrated in a small area for low-power conversion. Other conversion techniques such as voltage domain stacking and high-speed digital low-dropout regulators can also be used in combination to achieve higher conversion efficiency with small area overhead.
Fine-grained DVFS also require complex control of the whole chip for its robust and efficient operation, and I believe design automation assisted by more digitized control and modularized converters will greatly help in the design process of the whole power delivery network as well as in the performance optimization of load circuits.
Security and Reliability
As IoT devices become more prevalent, ensuring secure and reliable operation of every device will become more challenging. Also, as these IoT applications will handle more important and private data, there is a greater incentive and risk of these devices being hacked. Because it is difficult for the end-user to monitor each IoT device, the chance of physical attack on IoT devices will increase. Protection from these possible attacks, a side-channel attack through power supply lines for example, has not been considered seriously in many real applications.
On top of the system security issues, uncontrolled environment and reduced design margin for power and area efficiency makes the systems more vulnerable to PVT variations, and controlling these variations is another big issue that spans from transistors to high-level architectures. To mitigate these design challenges for security and reliability, we basically need more observability into on-chip states such as voltage, current and temperature, as well as more controllability over them. Low-cost sensors and detector logics in combination with high-efficiency power management blocks will greatly improve the protection over possible physical interventions and dynamic environmental changes.