MRI: Acquisition of a GPU Accelerated Vermont Advanced Computing Core

该项目将通过收购一个名为DeepGreen的高性能计算机集群来实现跨学科科学。基于尖端的大规模并行图形处理单元（GPU）技术，DeepGreen将被来自佛蒙特大学六个学院和整个东北部的300多名用户使用。这种独特的混合架构旨在优化人工智能（AI）应用，并允许在具有重大社会重要性的问题上取得快速进展。它们包括：量子计算、药物发现和设计、安全机器人、自适应作物害虫的控制，以及用于医疗保健和运输行业的新型计算机视觉工具。例如，DeepGreen将允许在世界上最大的非法吸毒者脑成像数据集上训练神经网络，从而产生对抗阿片类药物流行的新型健康和政策战略。科技团队的一个重点是扩大能够利用GPU硬件解决问题的人员数量，为当前和未来的人工智能经济提供所需的训练有素的多样化技术劳动力。DeepGreen是由一组来自物理、医学、生物、计算和农业科学的专家与一组经验丰富的信息技术专业人员合作设计的。它将能够基于最新的NVIDIA Tesla V100架构进行超过8千万亿次的混合精度计算，并采用混合设计，允许在异构计算节点之间传递高带宽消息。它的极端并行性将促进三个相互关联领域的研究：量子多体系统、分子模拟和建模、深度学习、人工智能和进化算法。深绿将打造变革性的研究管道。它将能够研究成千上万的量子纠缠原子，以及生物系统中数百万相互作用的组件，为结构-功能机制提供见解。机器学习和深度神经网络将利用DeepGreen的张量核来解决各种问题。这些问题包括：用于分子动力学模拟的粗粒度电位的开发，机器人众包决策的实时动态处理，入侵害虫的基因组测序，以及医学成像中的特征识别，以区分癌性肿瘤和良性结节。为DeepGreen设计的软件将作为开源软件向公众发布，其他科学家和研究人员可以立即使用和扩展它。该项目还将支持下一代数据科学家。以GPU计算和机器学习框架为重点的培训研讨会、新的大学课程，以及与现有的nsf资助的本地研究生培训计划的合作，将推动DeepGreen的广泛利用。该奖项反映了美国国家科学基金会的法定使命，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。

This project will enable interdisciplinary science through the acquisition of a high-performance computer cluster, named DeepGreen. Based on cutting-edge massively parallel graphics processing unit (GPU) technologies, DeepGreen will be utilized by the over 300 users from six Colleges at the University of Vermont, and throughout the Northeast. The unique hybrid architecture was designed to optimize artificial intelligence (AI) applications and will allow for rapid progress on problems of great societal importance. They include: quantum computing, drug discovery and design, safe robotics, control of adaptive crop pests, and new computer vision tools for use in the health care and transportation industries. As an example, DeepGreen will allow the training of neural networks on the world's largest brain imaging datasets of illicit drug users, yielding novel health and policy strategies to combat the opioid epidemic. A focus of the scientific and technical team is to broaden the number of personnel able to exploit GPU hardware for problem solving, producing the highly trained and diverse technical workforce required for the current and future AI economy. DeepGreen was designed by a team of experts from the physical, medical, biological, computational, and agricultural sciences, partnered with an experienced group of information technology professionals. It will be capable of over 8 petaflops of mixed precision calculations based on the latest NVIDIA Tesla V100 architecture with a hybrid design allowing high bandwidth message passing across heterogeneous compute nodes. Its extreme parallelism will facilitate research in three interconnected areas: quantum many-body systems, molecular simulation and modeling, and deep learning, artificial intelligence and evolutionary algorithms. DeepGreen will forge transformative research pipelines. It will enable the study of thousands of quantum entangled atoms, and millions of interacting components in biological systems providing insights into structure-function mechanisms. Machine learning and deep neural networks will exploit DeepGreen's Tensor Cores to solve diverse problems. These problems include: the development of coarse grained potentials for use in molecular dynamics simulations, real time dynamic processing of crowd sourced decision making for robotics, genomic sequencing of invasive pests, and feature recognition in medical imaging to distinguish cancerous tumors from benign nodules. Software designed for use on DeepGreen will be released to the public as open source, with other scientists and researchers being able to immediately use and extend it. This project will also support the next generation of data scientists. Training workshops focused on GPU computing and machine learning frameworks, new university courses, and partnerships with existing local NSF-funded graduate training initiatives, will drive broad utilization of DeepGreen.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.