CNS Core: Small: Re-engineering Applications for Tensor Processing Units

点击翻译按钮获取中文摘要

To overcome the inefficiency of conventional processors for machine learning (ML) and artificial intelligence (AI) computing applications, hardware accelerators that provide operators to compute on tensors/matrices directly emerge in all types of computer systems. Any application that naturally consumes and produces tensors/matrices may take advantage of these accelerators. However, these accelerators might only be used for ML/AI applications due to the lack of an appropriate programming interface and system support. This project will bridge the gap of making these accelerators available for more applications. This project will redesign applications to use these accelerators to improve performance and energy.This project will (1) build real systems using commercialized ML/AI accelerators (e.g., Google's Tensor Processing Units (TPUs)), (2) develop a programming interface that allows programmers to create any type of application on the built system, (3) redesign the algorithms of applications to use efficiently the operators that TPUs offer, (4) compose library functions and runtime systems to process efficiently tasks but hide complexity from programmers, and (5) revisit the design of the storage subsystem to supply data more efficiently. Through these aforementioned tasks, this project will demonstrate the challenges and analyze the potentials of extending the applications of TPUs. This project, if successful, will provide the first general-purpose programming platform for TPUs. By making the outcomes available to the public, this project will be available to all research areas that depend on computation/analytics of tensor/matrix datasets. In fact, the target applications in this project already cover database systems, bioinformatics algorithms and physics simulations. This project will also offer research/learning opportunities for a general audience, interdisciplinary researchers, and minority groups by classroom teaching, publications, talks and undergraduate summer interns.The research outcomes will be made through peer-reviewed conference and journal papers. The code developed and the configurations of hardware platforms will be publicly available through third party repository services (https://github.com/escalab/) after the research outcome is published. A cloud data storage service will be used to store all raw experimental data, copies of source code and applications datasets for at least three years and make them available upon appropriate requests.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.