CAREER: Generating Domain-Specific Systems at Scale

在技术扩展带来的回报不断减少的时代，硬件加速是为未来计算平台提供更高性能和能效的最有前途的方法之一。然而，最新的加速器研究主要集中在少数算法的单个加速器开发上，对如何在提高整体执行质量的同时整体生成多个加速器缺乏系统的了解。为了解决这一限制，该项目的目标是开发一个质量驱动的加速器设计流程，以支持端到端应用程序规模的特定领域加速器的整体模拟、设计和集成。它将特别有利于新兴的实时应用，如机器人和自动驾驶车辆，这些应用具有一系列计算要求高的内核。此外，综合教育活动将通过外联活动扩大各级对计算机的参与，使K-12、本科生和研究生参与硬件行业令人兴奋的机会。这个项目将促进计算机体系结构方面的知识，了解如何有效地大规模生成特定领域的系统，以实现端到端的加速。基于这一目标，本项目研究了以下三个协同研究方向：1)设计一个软硬件协同仿真基础设施，以定量评估闭环系统的端到端性能；2)开发一个整体设计空间探索框架，以高效地构建、表示和导航整个片上系统的复杂设计空间；以及3)通过软硬件协同设计，构建一个自适应的竞争管理系统，以动态划分加速器之间的共享资源。该项目开发的方法和框架将是开源的，并能够在比今天更大的规模上实现加速器集成，从根本上推动异构系统设计领域的发展。该奖项反映了NSF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。

In the era of diminishing returns from technology scaling, hardware acceleration is one of the most promising approaches to deliver improved performance and energy efficiency for future computing platforms. However, state-of-the-art accelerator research has largely focused on individual accelerator development for a handful of algorithms, without a systematic understanding of how to holistically generate many accelerators at scale while improving the overall execution quality. To address the limitation, the goal of this project is to develop a quality-driven accelerator design flow that enables the holistic simulation, design, and integration of domain-specific accelerators at scale for end-to-end applications. It will especially benefit emerging real-time applications, such as robotics and self-driving vehicles, which feature a range of computationally demanding kernels. In addition, the integrated education activities will broaden participation in computing at all levels with outreach activities to engage K-12, undergraduate, and graduate students in exciting opportunities in the hardware industry. This project will advance the state of knowledge in computer architecture on how to effectively generate domain-specific systems at scale to achieve end-to-end acceleration. Toward this goal, the project investigates the following three synergistic research thrusts: 1) design a hardware-software co-simulation infrastructure to quantitatively evaluate the end-to-end performance of closed-loop autonomous systems; 2) develop a holistic design-space exploration framework to efficiently construct, represent, and navigate the complex design space of entire systems-on-chip; and 3) build an adaptive contention management system through hardware-software co-design to dynamically partition shared resources between accelerators. The methodology and framework the project develops will be open-sourced and enable accelerator integration at a greater scale than what can be done today, fundamentally advancing the field of heterogeneous system 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.