An intellectual overlap of pure mathematics and engineering techniques targeted to develop self-reliant, efficient, and clean artificial intelligence processors

高耗能处理器被广泛应用于快速发展功能强大的云计算和人工智能(AI)。由于人们对深度学习的推崇，人工智能的能耗大幅增加。深度学习使用由数亿甚至数十亿个参数组成的非常大的数学模型来处理海量数据。它消耗大量能源，产生大量二氧化碳排放。根据《麻省理工学院技术评论》的数据，训练深度学习算法在包括制造在内的整个生命周期中排放的排放量几乎是普通美国汽车的五倍。一个训练下棋的人工智能系统可以产生19.2万磅的二氧化碳。如果我们不重新评估目前的人工智能研究议程，人工智能可能会在未来几年导致气候变化，并严重污染地球。这项提议的总体目标是通过优化程序的每个方面来降低人工智能数据中心(DC)的能源消耗。这些变化始于计算硬件和软件算法，并延伸到电源。我们的目标是让DC为智能数据收集、处理和学习应用程序做好准备。医疗保健和社会科学需要新的计算基础设施和DC来帮助专业人员提高这些关键方面的质量。我们需要结合和拓宽我们在电力电子、数学、深度学习、生物医学、社会科学和化学工程方面的知识和成就，为人工智能DC建设高效自力更生的基础设施。为了实现这一目标，我们将专注于：i)用清洁和可再生能源为AI DC提供动力；ii)优化算法和数据管理方法，以降低能源消耗；iii)为AI DC开发新的能源存储和电源管理基础设施。

Energy-consuming processors are widely used in rapidly developing powerful cloud-based computing and artificial intelligence (AI). AI's power consumption is enormously increased due to the rush in admiration of deep learning that processes huge amounts of data with very large mathematical models that can consist of hundreds of millions or even billions of parameters. It consumes a lot of energy and generates lots of CO2 emissions. According to MIT Technology Review, training deep-learning algorithms emits nearly five times as much emissions as the average American car throughout its entire life cycle, including manufacturing. An artificial intelligence system training to play chess can generate 192,000 pounds of CO2. If we do not reassess the current AI research agenda, AI could contribute to climate change in the years to come and significantly pollute the earth.The overarching goal of this proposal is to reduce the energy consumption of AI Data Centers (DCs) by optimizing every aspect of the procedure. The changes begin with computing hardware and software algorithms and extend to power sources. Our goal is to prepare DCs for smart data collection, processing, and learning applications. Healthcare and social sciences require new computing infrastructures and DCs to help professionals enhance the quality of these crucial aspects. We need to combine and broaden our knowledge and our achievements in power electronics, mathematics, deep learning, biomedical, social science, and chemical engineering to build high-efficiency self-reliant infrastructures for the AI DCs. To achieve this goal, we will focus on: i) powering AI DCs with clean and renewable energy sources; ii) optimizing the algorithms and data management methods to reduce energy consumption; iii) developing novel energy storage and power management infrastructure for AI DCs.