PHY Link Design and Optimization For High-Speed Low-Power Communication Systems

The ever-growing demands for high-bandwidth data transfer have been pushing towards advancing research efforts in the field of high-performing communication systems. Studies on the performance of single chip, e.g. faster multi-core processors and higher system memory capacity, have been explored. To further enhance the system performance, researches have been focused on the improvement of data-transfer bandwidth for chip-to-chip communication in the high-speed serial link. Many solutions have been addressed to overcome the bottleneck caused by the non-idealties such as bandwidth-limited electrical channel that connects two link devices and varieties of undesired noise in the communication systems. Nevertheless, with these solutions data have run into limitations of the timing margins for high-speed interfaces running at multiple gigabits per second data rates on low-cost Printed Circuit Board (PCB) material with constrained power budget. Therefore, the challenge in designing a physical layer (PHY) link for high-speed communication systems turns out to be power-efficient, reliable and cost-effective. In this context, this dissertation is intended to focus on architectural design, system-level and circuit-level verification of a PHY link as well as system performance optimization in respective of power, reliability and adaptability in high-speed communication systems. The PHY is mainly composed of clock data recovery (CDR), equalizers (EQs) and high- speed I/O drivers. Symmetrical structure of the PHY link is usually duplicated in both link devices for bidirectional data transmission. By introducing training mechanisms into high-speed communication systems, the timing in one link device is adaptively aligned to the timing condition specified in the other link device despite of different skews or induced jitter resulting from process, voltage and temperature (PVT) variations in the individual link. With reliable timing relationships among the interface signals provided, the total system bandwidth is dramatically improved. On the other hand, interface training offers high flexibility for reuse without further investigation on high demanding components involved in high costs. In the training mode, a CDR module is essential for reconstructing the transmitted bitstream to achieve the best data eye and to detect the edges of data stream in asynchronous systems or source-synchronous systems. Generally, the CDR works as a feedback control system that aligns its output clock to the center of the received data. In systems that contain multiple data links, the overall CDR power consumption increases linearly with the increase in number of links as one CDR is required for each link. Therefore, a power-efficient CDR plays a significant role in such systems with parallel links. Furthermore, a high performance CDR requires low jitter generation in spite of high input jitter. To minimize the trade-off between power consumption and CDR jitter, a novel CDR architecture is proposed by utilizing the proportional-integral (PI) controller and three times sampling scheme. Meanwhile, signal integrity (SI) becomes critical as the data rate exceeds several gigabits per second. Distorted data due to the non-idealties in systems are likely to reduce the signal quality aggressively and result in intolerable transmission errors in worst case scenarios, thus affect the system effective bandwidth. Hence, additional trainings such as transmitter (Tx) and receiver (Rx) EQ trainings for SI purpose are inserted into the interface training. Besides, a simplified system architecture with unsymmetrical placement of adaptive Rx and Tx EQs in a single link device is proposed and analyzed by using different coefficient adaptation algorithms. This architecture enables to reduce a large number of EQs through the training, especially in case of parallel links. Meanwhile, considerable power and chip area are saved. Finally, high-speed I/O driver against PVT variations is discussed. Critical issues such as overshoot and undershoot interfering with the data are primarily accompanied by impedance mismatch between the I/O driver and its transmitting channel. By applying PVT compensation technique I/O driver impedances can be effectively calibrated close to the target value. Different digital impedance calibration algorithms against PVT variations are implemented and compared for achieving fast calibration and low power requirements