CAREER: Rethinking PIM-Assisted GPU Computing for Multi-Tenant Artificial Intelligence

职业：重新思考用于多租户人工智能的 PIM 辅助 GPU 计算

• 批准号: 2239638
• 负责人: Chenchen Liu
• 金额: $ 53.9万
• 依托单位: University of Maryland Baltimore County
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2023
• 资助国家: 美国
• 起止时间: 2023-08-01 至 2028-07-31
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2239638&HistoricalAwards=false
• 关键词: CAREER; Rethinking; PIM; Assisted; GPU

Artificial intelligence (AI) systems have entered the “multi-tenant” era, where multiple deep neural network (DNN) models are executed simultaneously. This involves concurrent deployment, computation, and interaction of multiple DNN models, increasing computational complexity and triggering new challenges: (1) How can a scalable and flexible computing architecture be realized that can adaptively host heterogeneous and concurrent DNN models? (2) How can computing flexibility requirements in multi-tenant DNN scenarios be met? (3) How can an efficient, end-to-end toolchain for building next-generation AI applications be realized in this context? This project addresses these challenges through three research thrusts: Thrust 1 investigates a novel processing-in-memory (PIM)-assisted graphics processing unit (GPU) architecture with innovative multi-tenant support, addressing important resource contention and model interaction issues. Thrust 2 explores dedicated GPU- and PIM-oriented scheduling techniques to enhance the platform’s performance. Finally, thrust 3 further enhances the multi-tenant AI application development cycle with algorithm optimization and code deployment support.With the successful completion of these thrusts, this project can achieve breakthroughs in modern AI computing and support the next generation of AI applications. The proposed techniques have the potential to accelerate AI design and deployment, spurring even wider AI utilization. This can contribute to important application areas with societal importance, including autonomous driving, metaverse immersion, smart agriculture, and industrial infrastructure. This project will also benefit students --and by consequence, society-- by incorporating research results within relevant courses, increasing the participation of women and other underrepresented groups in computing, and sharing research results with researchers, companies, and government agencies.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.

人工智能（AI）系统已经进入了“多租户”时代，多个深度神经网络（DNN）模型同时执行。这涉及到多个DNN模型的并发部署、计算和交互，增加了计算复杂性并引发了新的挑战：（1）如何实现可扩展且灵活的计算架构，以自适应地托管异构和并发DNN模型？(2)如何满足多租户DNN场景中的计算灵活性要求？(3)在这种情况下，如何实现用于构建下一代AI应用程序的高效端到端工具链？该项目通过三个研究重点来解决这些挑战：重点1研究了一种新颖的内存处理（PIM）辅助图形处理单元（GPU）架构，具有创新的多租户支持，解决了重要的资源争用和模型交互问题。Thrust 2探索了专用的面向GPU和PIM的调度技术，以增强平台的性能。最后，thrust 3通过算法优化和代码部署支持，进一步提升多租户AI应用开发周期，随着这些thrust的顺利完成，本项目可以实现现代AI计算的突破，支撑下一代AI应用。所提出的技术有可能加速人工智能的设计和部署，从而促进更广泛的人工智能利用。这可以为具有社会重要性的重要应用领域做出贡献，包括自动驾驶、虚拟实境沉浸、智能农业和工业基础设施。该项目还将使学生受益，从而使社会受益，将研究成果纳入相关课程，增加妇女和其他代表性不足的群体对计算的参与，并与研究人员、公司、该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响进行评估，被认为值得支持审查标准。