Ultra-low-voltage (ULV) operation where the supply voltage of the digital computing hardware is scaled down to the level near or below transistor threshold voltage (e.g. 300-500mV) is a key technique to achieve high computing energy efficiency. It has enabled many new exciting applications in the field of Internet of Things (IoT) devices and energy-constrained applications such as medical implants, environment sensors, and micro-robots. Ultra-low-voltage (ULV) operation is also commonly used with the emerging architectures that are often non Von-Neumann style to empower energy-efficient cognitive computing.
One the biggest challenge in realizing ULV design is the large circuit delay variability. To guarantee functionality in the worst-case process, voltage, and temperature (PVT) condition, the traditional safety margin approach requires operating at a slower clock frequency or higher supply voltage which significantly limits the achievable energy efficiency of the hardware. To fully claim the energy efficiency of ULV, the large circuit delay variation needs to be adaptively handled. However, the existing adaptive techniques that are optimized for nominal supply voltage operation and traditional Von-Neumann architectures become inefficient for ULV designs and emerging architectures.
This thesis presents adaptive techniques based on timing error detection and correction (EDAC) that are more suitable for the energy-constrained ULV designs and the emerging architectures. The proposed techniques are demonstrated in three test chips: (1) R-Processor: A 0.4V resilient processor with a voltage-scalable and low-overhead in-situ EDAC technique. It achieves 38% energy efficiency improvement or 2.3X throughput improvement as compared to the traditional safety margin approach. (2) A 450mV timing-margin-free waveform sorter for brain computer interface (BCI) microsystem. It achieves 49.3% higher energy efficiency and 35.6% higher throughput than the traditional safety margin approach. (3) Ultra-low-power and robust power-management system which consists of a microprocessor employing ULV EDAC, 63-ratio integrated switched-capacitor DC-DC converter, and a fully-digital error based regulation controller.
In this thesis, we also explore circuits for emerging techniques. The first is temperature sensors for dynamic-thermal-management (DTM). The modern high-performance microprocessors suffer from ever-increasing power densities which has led to reliability concerns and increased cooling costs from excessive heat. In order to monitor and manage the thermal behavior, DTM techniques embed multiple temperature sensors and use its information. The size, accuracy, and voltage-scalability of the sensor are critical for the performance of DTM. Therefore, we propose a temperature sensor that directly senses transistor threshold voltage and the test chip demonstrates 9X smaller area, 3X higher accuracy, and 200mV lower voltage scalability (down to 400mV) than the previous state-of-art.
Another area of exploration is interconnect design for ultra-dynamic-voltage-scaling (UDVS) systems. UDVS has been proposed for applications that require both high performance and high energy efficiency. UDVS can provide peak performance with nominal supply voltage when work load is high. When work load is moderate or low, UDVS systems can switch to ULV operation for higher energy efficiency. One of the critical challenges for developing UDVS systems is the inflexibility in various circuit fabrics that are often optimized for a single supply voltage. One critical example is conventional repeater based long interconnects which suffers from non-optimal performance and energy efficiency in UDVS systems. Therefore, in this thesis, we propose a reconfigurable interconnect design based on regenerators and demonstrate near optimal performance and energy efficiency across the supply voltage of 0.3V and 1V.
Identifer | oai:union.ndltd.org:columbia.edu/oai:academiccommons.columbia.edu:10.7916/D8TM7BPF |
Date | January 2016 |
Creators | Kim, Seongjong |
Source Sets | Columbia University |
Language | English |
Detected Language | English |
Type | Theses |
Page generated in 0.0029 seconds