深度学习由于存在需要海量标注数据、通用智能水平弱、高度依靠计算力等局限，难以成为解决人工智能问题的终极手段。各国纷纷启动脑计划，希望通过发展类脑计算，克服深度学习的不足。类脑计算的基础是脉冲神经网络（SNN, Spiking Neural Network），因此类脑研究的重要任务是研制适合SNN的高能效类脑计算机。通过工作负载表征对SNN负载精确“画像”对类脑计算架构的研究至关重要。本课题拟1）充分利用大规模分布式FPGA异构平台在横向（节点规模可伸缩）及纵向（软硬件比例可调整）的扩展能力，对SNN的负载特性进行系统性的量化实验；2）对负载特性数据进行理论分析并建立基准测试模型并设计相应基准测试程序集；3）运用目标优化及聚类算法解决SNN模型与类脑平台自动映射问题，提出基准测试程序与硬件平台的自动映射方法，加速基准测试进度。课题的工作将为类脑体系结构研究提供理论依据与基准测试支撑。

Deep learning is difficult to be the ultimate means to solve AI problems because of its limitations such as requirement for large amount of labeled data, weak general intelligence level and high dependence on computational power. Many countries have launched brain programs, hoping to overcome the shortcomings of deep learning by developing brain-like computing. Brain-like computing is based on spiking neural network (SNN), so the important task of brain-like research is to develop a high-efficiency brain-like computer suitable for SNN. The accurate "portrait" of SNN workload through workload characterization is crucial to the research of brain-like computing architecture. This project is to make full use of the scalability of large-scale distributed heterogeneous platform of FPGA in horizontal (scalable node size) and vertical (adjustable hardware and software ratio) to systematically quantify the workload characteristics of SNN; 2) to theoretically analyze the workload characteristics and establish a benchmark model and design the corresponding benchmark code; 3) to optimize and cluster the target. The algorithm solves the problem of automatic mapping between SNN model and brain-like platform, and proposes an automatic mapping method between program and hardware platform to accelerate benchmarking progress. This project will provide theoretical basis and benchmark support for the study of brain-like architecture.