基于Nor Flash的存算一体人工智能芯片和系统研究

随着人工智能技术在当前社会的快速普及，世界各国都在大力推进物联网、智慧城市等相关新兴产业的建设，智能万物互联时代已经势不可挡。但是，随着神经网络计算参数量的不断增加，数据传输过程所引起的能量和时间消耗已经远超数据计算过程，成为限制处理器性能提升的关键因素。人们意识到传统采用存算分离的冯·诺依曼架构成为主要制约因素，已不适合大数据驱动的人工智能应用需求，需要采用变革性新型计算模式。基于生物脑启发的存算一体架构解决了传统微处理器中由于频繁数据传输所带来的时间和能量消耗的问题，可以从根本上解决冯·诺依曼瓶颈所引发的能效问题。本项目瞄准学术与产业前沿，从器件、架构、软件等层面进行创新，研究具有自主知识产权的、基于新型Nor Flash介质的存算一体人工智能芯片系统，能够大幅减少数据搬运，极大提高计算并行度和能效，突破冯·诺依曼架构算力与功耗瓶颈，为人工智能应用提供芯片技术支撑，推动产业升级。

With the incoming of an artificial intelligence (AI) era, neural networks are greatly revolutionizing the human society. Many countries have vigorously promoted the development of the Internet of things (IoT), smart cities and other emerging industries. The era of intelligent Internet of everything has become overwhelming. However, with the increasing number of neural network computing parameters, the separation of the computing unit and memory unit in the traditional von-Neumann computing architecture also hinders the further performance improvement of current information processing systems. Therefore, innovative computing paradigms and system architectures have become a necessity. This can significantly reduce the high energy and time consumption caused by the frequent movement of data between the storage unit and the computing unit during the computing process. Brain-spired computing-in-memory (CIM) architecture can solve the delay and energy consumption caused by frequent data transmission in traditional computing architecture, and can fundamentally solve the energy efficiency problem caused by von Neumann bottleneck. This project focuses the novel CIM technique. A full study, including the device, circuit, architecture, software and algorithm, will be fulfilled. Based on the new type of NOR Flash bit-cell process chip system, an CIM chip will be implemented for cloud/terminal artificial intelligence application. The chip can reduce the data handling and improve energy efficiency. The project will support the AI chip application support and promote related industrial upgrading.