Exploiting Hardware Heterogeneity for Efficient Execution of Emerging Applications

点击翻译按钮获取中文摘要

The proposed work will develop computing systems that will address the urgent need for performance and energy efficiency. This research will impact future energy consumption, energy costs, and greenhouse gas emissions, and will consequently lead to significant economic and environmental benefits for Canadians. The major emphasis of this work is on building future power and energy-efficient systems that can significantly improve the efficiency of execution for many important applications in of computer vision, artificial intelligence, bioinformatics, and web search. Our research aims to achieve this energy efficiency by using heterogeneous computational systems with different types of hardware-based accelerators. The key uniqueness of this approach is both the breadth and depth of techniques we will apply across the whole systems stack: hardware-based acceleration, data compression, and in-memory computation. In most geographical locations (including Canada) there is shortage of highly qualified personnel (HQPs) in the computer architecture and systems fields. This program will train many HQPs in computing systems who will gain important skills in computer architecture, systems, performance analysis, compilers, low-level and systems programming, and programming existing hardware accelerators, such as graphics processing units (GPUs) and field-programmable gate-arrays (FPGAs). ******This program consists of several major research directions.******1. We aim to identify the right programming models and tuning/mapping algorithms for the most efficient use of existing hardware accelerators. Currently, the burden of doing so falls almost entirely on individual application developers. ******2. We will search for efficient hardware and software optimizations to improve the resource and memory use of modern hardware accelerators in emerging applications such as deep neural network (DNNs). In particular, DNN training is a process that requires large amounts of memory that are not available even in state-of-the-art GPU accelerators. We will leverage our expertise in energy- and performance-efficient data compression and encoding to significantly reduce the memory consumption of existing DNN training algorithms. ******3. We will explore the potential of more radical optimizations that would include (but are not limited to) (1) lossy compression and (2) direct execution on compressed data, which would allow to improve both performance and energy efficiency even further.