high performance computing on graphics processing units: hgpu.org

The confluence of the rapid growth in electronic data in recent years, and the renewed interest in domain-specific hardware accelerators presents exciting technical opportunities. Traditional scale-out solutions for processing the vast amounts of text data have been shown to be energy- and cost-inefficient. In contrast, custom hardware accelerators can provide higher throughputs, lower latencies, and significant energy savings. In this thesis, I present a set of hardware accelerators for unstructured big-data processing and natural language processing. The first accelerator, called HAWK, aims to speed up the processing of ad hoc queries against large in-memory logs. HAWK is motivated by the observation that traditional software-based tools for processing large text corpora use memory bandwidth inefficiently due to software overheads, and, thus, fall far short of peak scan rates possible on modern memory systems. HAWK is designed to process data at a constant rate of 32 GB/s-faster than most extant memory systems. I demonstrate that HAWK outperforms state-of-the-art software solutions for text processing, almost by an order of magnitude in many cases. HAWK occupies an area of 45 sq-mm in its pareto-optimal configuration and consumes 22 W of power, well within the area and power envelopes of modern CPU chips. The second accelerator I propose aims to speed up similarity measurement calculations for semantic search in the natural language processing space. By leveraging the latency hiding concepts of multi-threading and simple scheduling mechanisms, my design maximizes functional unit utilization. This similarity measurement accelerator provides speedups of 36x-42x over optimized software running on server-class cores, while requiring 56x-58x lower energy, and only 1.3% of the area.