深度学习被公认为是最重要的智能处理技术之一，是超级计算机、数据中心、智能手机到嵌入式设备等各种不同应用场景上的重要负载。然而，现有各种芯片难以同时满足各种场景对深度学习处理高性能、低能耗、低延迟的三大需求。为研制能满足上述三大需求的新型深度学习处理器，本项目拟充分利用深度学习本身对于计算过程中非精确性的容忍能力，通过适当放松计算的精确度，提升芯片的深度学习处理能力。我们拟抽象出非精确深度学习处理器的指令集（含计算、访存和控制指令），并面向该指令集探索深度学习处理器的微结构、算法库和系统软件。上述研究成果将被集成到一款深度学习处理器的样片中。我们还将基于该样片构建拟面向健康管理的有界误差深度学习应用示范。通过指令集、微结构、算法库、系统软件以及示范性应用五方面全链式的研究，本项目有望从结构层面上实现深度学习处理器速度和能效的数量级提升，为国产深度学习处理器的发展提供借鉴。

Deep learning is widely recognized as one of the most important intelligent processing techniques, and it is crucial tasks on supercomputers, data centers, mobile phones, and embedded devices. However, current chips cannot simultaneously fulfill the three requirements of deep learning applications: high performance, low energy-consumption, and low latency. To develop a novel deep learning architecture which can fulfill the three requirements, this project plans to leverage the inaccuracy tolerance ability of deep learning, so as to promote the efficiency of deep learning chip. We plan to extract an instruction set (including computation instructions, memory access instructions, and control instructions) for inaccurate deep learning processing. And based on the instruction set, we will investigate the microarchitecture, library, and system software of deep learning processor. The above investigations will be adopted by a test chip of deep learning processor. With this test chip, we will also build a series of inaccurate deep learning application demo about healthy management. Our systematically research on deep learning processor may promote the performance and energy-efficiency of deep learning processors by two orders of magnitude, and provide helpful suggestions to the development of domestic deep learning processors.