Collaborative Research: CNS Core: Small: NV-RGRA: Non-Volatile Nano-Second Right-Grained Reconfigurable Architecture for Data-Intensive Machine Learning and Graph Computing

在数字数据时代，计算设备随着时间的推移不断积累海量数据。这导致了计算模式的转变，采用了机器学习(ML)和图形分析处理来分析如此海量的数据。传统的计算模式在能耗、延迟和计算效率方面效率低下。现有的内存计算范例在一定程度上解决了这一挑战，但不适用于异类应用程序。该项目引入了一种新颖的计算体系结构--右粒度可重构体系结构(RGRA)，它结合了粗粒度可重构阵列(CGRA)的灵活性和FPGA的可编程性，采用了具有环状拓扑结构的电路交换互连和路由器网络来应对这一挑战。在顶层，RGRA是一个多核体系结构，每个核都可以更细的粒度进行配置。拟议的研究还有望导致开发新的可重构体系结构分支，这些可重构体系结构支持图分析等数据密集型应用的高效执行，并受益于可重构性和异构性等体系结构方面。就更广泛的影响而言，硬件加速器的设计是计算机体系结构领域的推动方向之一。因此，无论应用程序的内存特性如何，RGRA的成功设计都是性能高效的，可能会产生重大的社会和经济影响。例如，它可以增强现有和新兴系统中的CPU、FPGA和GPU。该项目的成果将包括高速可重新配置的NVM和互连的设计，这也可以被用于开发高吞吐量处理器的多核系统。由于ML在高中授课，该项目有很好的扩展到社区和吸引学生的机会，特别是在以下方面：(I)招募代表不足的班级，包括少数族裔和女性；(Ii)以K-12的形式扩展，并通过暑期实习和高级设计项目让本科生参与研究；以及(Iii)提供关于ML加速器设计的研究生课程。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 digital data, computing devices amass vast amounts of data continuously over time. This leads to a paradigm shift in the computing by adopting machine learning (ML) and graph analytic processing to analyze such massive amounts of data. Traditional computing paradigms are inefficient in terms of energy consumption, latency, and computational efficiency. Existing in-memory computing paradigms address this challenge to a certain extent, but are not adaptable for heterogeneous applications. The proposed project introduces a novel computing architecture, right-grained reconfigurable architecture (RGRA) that combines the flexibility of coarse-grained reconfigurable array (CGRA) and programmability of FPGAs, deploying a circuit-switched interconnects and router network with torus topology to address this challenge. At the top-level, RGRA is a many-core architecture with each core configurable at finer granularity. The proposed research is also expected to lead to the development of new branch of reconfigurable architectures that support efficient execution of data-intensive applications such as graph analytics and benefit from architectural aspects such as reconfigurability and heterogeneity. In terms of broader impact, design of hardware accelerators is one of the driving directions in the field of computer architecture. As such, the successful design of RGRA that is performance efficient irrespective of the application memory-traits can have a significant societal and economic impact. For instance, it can augment CPUs, FPGAs, and GPUs in the existing and emerging systems. The results of the project will include design of high-speed reconfigurable NVMs and interconnects, which can also be adopted in many-core systems towards developing high-throughput processors. With ML being taught in the higher-secondary schools, the project has a good scope for outreach to the community and attract students, especially in terms of (i) recruitment of underrepresented classes including minorities and women; (ii) outreach in the form of K-12 and undergraduate student involvement in research via summer internships and senior-design projects; and (iii) offering a graduate course on ML accelerator 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.