Elements: Software Infrastructure for Programming and Architectural Exploration of Neuromorphic Computing Systems

事实证明，机器学习在医疗保健、环境、教育、基础设施和网络安全等一系列社交领域都取得了巨大成功。目前用于运行机器学习任务的计算平台有很高的碳足迹。模仿生物神经元和突触的神经形态计算系统可以以高度节能的方式执行这些任务。然而，神经形态计算的主要挑战在于它被用户采用，从系统开发商的角度来看，以应对新神经形态芯片设计更快的上市压力。该项目开发了名为NeuroXplorer的软件基础设施，可帮助最终用户和神经形态系统的开发人员：它允许以尽可能高效的方式将机器学习任务映射到神经形态芯片上；并提供分析、模拟和合成工具，可用于探索新的芯片设计，以满足新兴的机器学习工作负载的需求。该项目的智力优势在于在NeuroXplorer内部开发编译器后端，以从机器学习任务的高级规范生成神经形态芯片的可执行代码；开发映射和合成工具，以在使用现场可编程门阵列(现场可编程门阵列)构建的新神经形态架构上执行机器学习任务；以及开发高性能软件，用于新神经形态架构的硬件/软件设计空间探索。NeuroXplorer是为模块化和可扩展而构建的，这样开发人员就可以轻松地为软件贡献新功能。可以通过互联网访问NeuroXplorer的功能。最终用户使用标准工作流训练机器学习模型并上传，在此基础上自动生成适当的代码并在神经形态体系结构上执行。最终的现场可编程门阵列设计的神经形态程序和位流文件可以免费下载。NeuroXplorer中的设计空间探索工具有效地解决了神经形态系统日益复杂的问题，以及将新兴设计技术整合到这些系统中的挑战。从教育的角度来看，该项目让德雷克塞尔大学的研究生和本科生参与软件的开发。来自学术界和工业界的合作者提供关于神经形态硬件、系统软件和应用程序的当前发展的客座讲座，这些讲座被整合到相关课程中。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。

Machine learning has proved to be immensely successful across a range of social domains such as healthcare, environment, education, infrastructure, and cybersecurity. Computing platforms currently used to run machine-learning tasks have a high carbon footprint associated with them. Neuromorphic computing systems, which mimic biological neurons and synapses can implement these tasks in a highly energy-efficient fashion. Major challenges for neuromorphic computing, however, lie in its adoption by users and from a system developer's perspective, to cope with faster time-to-market pressure for new neuromorphic chip designs. This project develops a software infrastructure called NeuroXplorer, which helps both end-users as well as developers of neuromorphic systems: it allows for machine-learning tasks to be mapped onto neuromorphic chips in the most efficient way possible; and provides analysis, simulation, and synthesis tools that can be used to explore new chip designs to meet the needs of emerging machine-learning workloads. NeuroXplorer is distributed under an open-source license to promote the adoption of neuromorphic computing as well as the development and commercialization of neuromorphic systems in the United States.The intellectual merits of the project lie in the development of compiler backends within NeuroXplorer to generate executable code for neuromorphic chips such as Loihi, Dynamic Neurormorphic Asynchronous Processor, and Microbrain from a high-level specification of the machine-learning task; development of mapping and synthesis tools to execute machine-learning tasks on novel neuromorphic architectures built using Field-Programmable Gate Array (FPGA); and development of high-performance software for hardware/software design-space exploration of new neuromorphic architectures. NeuroXplorer is built to be modular and extensible such that developers can easily contribute new features to the software. The capabilities of NeuroXplorer are accessible over the Internet. The end-user trains the machine-learning model using a standard workflow and uploads it, upon which the appropriate code is automatically generated and executed on neuromorphic architecture. The neuromorphic program and bitstream files for the final FPGA design can be freely downloaded. Design-space exploration tools within NeuroXplorer efficiently tackle the growing complexity of neuromorphic systems and challenges in integrating emerging design technologies into these systems. From an educational perspective, the project involves both graduate and undergraduate students at Drexel University in the development of the software. Collaborators from academia and industry deliver guest lectures on current developments in neuromorphic hardware, system software, and applications, with these lectures being integrated within relevant courses.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.