CRII:CSR: Enabling High-Performance Deep Learning Computing System via Software and Hardware Co-Optimized Reconfiguration

CRII：CSR：通过软硬件协同优化重构实现高性能深度学习计算系统

• 批准号: 1939380
• 负责人: Chenchen Liu
• 金额: $ 17.47万
• 依托单位: University of Maryland Baltimore County
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2019
• 资助国家: 美国
• 起止时间: 2019-07-01 至 2023-06-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1939380&HistoricalAwards=false
• 关键词: CRII; CSR; Enabling; Performance; Deep

Deep learning (DL) technology has much success in today's artificial intelligence (AI) applications. However, executions of DL algorithms consume much computational and energy resources because of the large-scale data and complex computation models. The project addresses this challenge with a hardware and software co-optimization strategy. A high-performance deep learning system is enabled with the capability to select automatically the best parameter configurations of advanced algorithms based on the hardware architecture of the computing system. The project seeks to solve a fundamental challenge of a DL based computing system where intensive computations are introduced while the computing resources and the real-time budget are limited. Task 1 proposes an algorithm-driven DL computing system configuration based on a full parameter deep learning compression and a hardware-friendly algorithm deployment. Task 2 investigates an architecture-driven DL computing system configuration based on DL computation performance profiling and modeling on various hardware architectures. The success of the project paves the design foundation of a DL based intelligence system by considering the constraints of data, algorithm, and hardware platform.The project provides benefit to computational intelligence (CI), embedded systems, mobile intelligence, machine learning, and computing architecture. The project will further promote software and hardware co-development towards highly efficient intelligence systems for real-world artificial intelligence applications. The project will also benefit a wide range of communities by means of seminar broadcasts, a short course program, and children and high-school programs. The education plan will enhance existing curricula and pedagogy by integrating interdisciplinary modules with innovative teaching practices. The outcomes of the project, including data and experiments results, will be distributed in the form of journal articles, conference proceedings, workshops, invited presentations and student thesis. The results related with developed curriculum will also be published in appropriate education conferences. The simulators built for validation and simulation code will also be available to the public. Copies of these papers, presentations, models, simulation codes and course notes will be placed on a research page (http://if-lab.org/awards/crii2018) associated with this project.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.

深度学习技术在当今的人工智能应用中取得了很大的成功。然而，由于海量数据和复杂的计算模型，执行DL算法消耗了大量的计算和能量资源。该项目通过硬件和软件共同优化战略来应对这一挑战。使高性能深度学习系统能够基于计算系统的硬件体系结构自动选择高级算法的最佳参数配置。该项目旨在解决基于数字图书馆的计算系统的一个基本挑战，在该系统中引入密集的计算，而计算资源和实时预算是有限的。任务1提出了一种基于全参数深度学习压缩和硬件友好算法部署的算法驱动的下行计算系统配置。任务2基于不同硬件体系结构上的DL计算性能分析和建模，研究了体系结构驱动的DL计算系统配置。该项目的成功为基于数据、算法和硬件平台的智能系统的设计奠定了基础，为计算智能(CI)、嵌入式系统、移动智能、机器学习和计算体系结构提供了有益的借鉴。该项目将进一步促进软件和硬件的共同开发，朝着用于现实世界人工智能应用的高效智能系统发展。该项目还将通过研讨会广播、短期课程计划以及儿童和高中计划等方式使广泛的社区受益。教育计划将通过将跨学科模块与创新的教学实践相结合，加强现有课程和教学方法。该项目的成果，包括数据和实验结果，将以期刊文章、会议记录、研讨会、特邀演讲和学生论文的形式分发。与制定的课程相关的结果也将在适当的教育会议上公布。为验证和模拟代码而构建的模拟器也将向公众开放。这些论文、演示文稿、模型、模拟代码和课程笔记的副本将放置在与此项目相关的研究页面(http://if-lab.org/awards/crii2018))上。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

项目成果

Automated Runtime-Aware Scheduling for Multi-Tenant DNN Inference on GPU

• DOI: 10.1109/iccad51958.2021.9643501
• 发表时间: 2021-11
• 期刊: 2021 IEEE/ACM International Conference On Computer Aided Design (ICCAD)
• 作者: Fuxun Yu;Shawn Bray;Di Wang;Longfei Shangguan;Xulong Tang;Chenchen Liu;Xiang Chen

Powering Multi-Task Federated Learning with Competitive GPU Resource Sharing

• DOI: 10.1145/3487553.3524859
• 发表时间: 2022-04
• 期刊: Companion Proceedings of the Web Conference 2022
• 作者: Yongbo Yu;Fuxun Yu;Zirui Xu

Enabling efficient ReRAM-based neural network computing via crossbar structure adaptive optimization

• DOI: 10.1145/3370748.3406581
• 发表时间: 2020-08
• 期刊: Proceedings of the ACM/IEEE International Symposium on Low Power Electronics and Design
• 作者: Chenchen Liu;Fuxun Yu;Zhuwei Qin;Xiang Chen