New Foundations for Next-Generation Reliable Throughput Architecture Design

Given the extraordinary computational power, general-purpose computing on graphics processor units (GPGPUs) emerge as a highly attractive platform for a wide range of HPC applications which have strong data-level or thread-level parallelism. However, current GPUs have limited capability in error detection and fault tolerance because they were originally designed for graphics processing applications, which are inherently fault tolerant. Many HPC applications, in contrast, require strict execution correctness. It is imperative to develop a novel set of error characterization and optimization techniques for GPUs. The goal of this research project is to construct new foundations for reliability characterization and prediction, error detection and fault tolerance in next-generation throughput processors (e.g., GPGPUs).

Enabling Energy-Efficient Computing in Mobile Platforms

Mobile devices, such as those found in smart phone and tablet devices, are quickly becoming the most widely used processors in consumer devices. Since their major power supply is battery, the low-power yet high-performance computing is highly desired by consumers. The goal of this research project is to explore energy-efficient computing in mobile platforms.

Combating the Power Challenge in GPGPUs

Current and future GPGPUs confront power and energy as the dominant constraint. The number of transistors integrated on a single GPU chip continues to increase due to the shrink of feature size and the demand for massively parallel computing cores to increase the throughputs. On the other hand, the continuous decrease of transistor supply voltage at each new technology node has largely stalled because of leakage constraints, leading to an ever-increasing power density. Therefore, GPGPUs must become more inherently energy efficient to resist the power wall. The objective of this research project is to explore a synergetic program to holistically and hierarchically improve the GPGPUs energy efficiency.