POSE: Phase II: Building an Open-Source Ecosystem for Deep-Learning Hardware-Software Co-Design

特定领域的加速是进一步提高性能和能源效率的最有前途的方法之一，如深度学习。尽管有大量的初创公司和大公司为深度学习开发专门的硬件和软件，但所有现有的实现都是专有的，没有一个可行的、自由开源的深度学习硬件-软件堆栈。由于缺乏开放共享的生态系统，不仅难以比较不同的实现方式，还大大减缓了创新速度，增加了设计成本，因为每个组织都需要从头开始实施。该提案的总体目标是建立一个开源生态系统，使全栈深度学习系统能够大规模开发，以构建下一代深度学习平台。如果成功的话，这个项目的成果将指向一个未来，在深度学习硬件或软件方面有伟大想法的开发人员可以快速评估，设计，并在真实的硬件和软件上以端到端的方式展示他们的想法，从而大大降低设计成本并加快创新步伐。具体来说，开源产品由三个基于我们成熟研究项目的关键组件组成：用于代码生成的Exo（https：//github.com/exo-lang/exo），用于深度学习加速器设计的Gemmini（https：//github.com/ucb-bar/gemmini）和用于片上系统集成的Chipyard（https：//github.com/ucb-bar/chipyard）。特别是，拟议的硬件-软件生态系统将从根本上解决以下挑战：1）如何使用软件堆栈评估深度学习加速器的端到端性能; 2）如何为特定场景生成高效的深度学习硬件加速器;以及3）如何将加速器与通用核心集成并在FPGA和制造中对其进行评估。该项目建立的软硬件生态系统将为未来通过软硬件协同设计实现深度学习加速的发展创建一个多元化和高度积极的开源社区。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审查标准进行评估，被认为值得支持。

Domain-specific acceleration is one of the most promising approaches to further improve performance and energy efficiency applications like deep learning. Despite the large number of startups and large companies developing specialized hardware and software for deep learning, all the existing implementations are proprietary, without a viable, freely open-source deep-learning hardware-software stack. The lack of an open and shared ecosystem not only makes it extremely hard to compare different implementations, it also significantly slows innovation and increases design costs as every organization needs to start their implementations from scratch.The overall objective of this proposal is to establish an open-source ecosystem that enables the development of full-stack deep-learning systems at scale to build next-generation deep-learning platforms. If successful, the outcomes of this project point toward a future in which developers with a great idea in either deep-learning hardware or software can quickly evaluate, design, and demonstrate their idea in an end-to-end fashion on real hardware and software, significantly lowering the design cost and accelerating the pace of innovation. Specifically, the open-source product consists of three key components based on our mature research projects: Exo for code generation (https://github.com/exo-lang/exo), Gemmini for deep-learning accelerator design (https://github.com/ucb-bar/gemmini), and Chipyard for system-on-chip integration (https://github.com/ucb-bar/chipyard). In particular, the proposed hardware-software ecosystem will fundamentally address challenges in 1) how to evaluate the end-to-end performance of deep-learning accelerators with the software stack; 2) how to generate efficient deep-learning hardware accelerators for specific scenarios; and 3) how to integrate accelerators with general-purpose cores and evaluate them in FPGA and fabrication. The hardware-software ecosystem established from this project will create a diverse and highly motivated open-source community for future developments on deep-learning acceleration through hardware-software co-design.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.