CAREER: Effectuating Hardware-Accelerated Systems with Domain-Specific Primitives

点击翻译按钮获取中文摘要

With enormous volumes and vast varieties of data being produced and consumed every second, having computer systems that can process large volumes and diverse sets of data in unprecedented efficiency is imperative for computer architects and system designers to keep up with this data deluge. Custom hardware has long been a solution to run a specific application or a set of applications at exceptional performance and energy efficiency compared to general purpose platforms. However, the development and deployment of custom hardware has been time-consuming and expensive. Moreover, the toolchain, programming, and usage of custom hardware are complex and obtuse, reducing its accessibility. Thus, there is a need to innovate and drastically reduce the development and deployment costs of custom hardware while making these accelerated systems accessible to not only computer system designers but also non-hardware savvy scientists to expedite their research analyses and advance knowledge in their perspective fields. This research develops a novel execution paradigm to democratize hardware accelerated system designs. This research approach leverages application-specific knowledge to guide the underlying hardware design and compose highly efficient, programmable, portable, and easy to use custom hardware systems that are orders of magnitude faster and more power efficient. This research promotes the progress of computer science and computer system designs and serves as an education tool in the computer research community to advance knowledge in hardware acceleration abstractions, concepts, and effectuations. In addition, this research can be applied to several example domains to solve open problems in the interdisciplinary area of computer architecture and healthcare such as cancer research and drug discovery to improve national health and bring direct impacts to human flourishing. This project explores how to create highly efficient hardware-accelerated systems for compute- and memory-intensive applications that process enormous volumes of data. The key idea is to develop a new execution paradigm, that decomposes algorithms into coarse-grained operators instead of traditional execution work units such as instructions, that describes how and what portions of applications can be accelerated and executed on a hardware-accelerated system. Such execution work units are called Domain-Specific Primitives (DSPs). The vision is to raise the level of abstraction when designing hardware accelerators by examining domain-specific, already-defined software datatypes and constructs and thereby creating hardware execution units that closely match software datatype-specific method calls, the starting point of DSPs. DSPs can be used by software compilers, programmers, as well as hardware development designers to compose, program, develop, and optimize domain-specific hardware-accelerated systems in a systematic fashion. This project first develops the fundamental concepts and abstractions of DSP designs. This project then explores the productivity, resultant system performance and energy efficiency, portability, and the ease of use of DSP-composed accelerated systems. The result is a transformative realization in highly performant, highly efficient, and highly productive development of hardware-accelerated systems that are easy to use. This research makes a significant contribution towards teaching acceleration concepts via various education and outreach programs. In addition, the research will advance computer systems and healthcare research and enable industry technology transfer.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.