一款人工智能芯片上FCOS模型的应用研究

Application of FCOS Model on an AI Chip

人工智能芯片是专门用于高效执行人工智能计算任务的芯片。中国电子科技集团公司第三十八所研制了一款针对边缘侧深度学习模型推理计算的人工智能芯片,主要面向雷达图像目标识别、色选机图像智能处理等应用。该芯片是一个异构的SOC芯片,由中央处理核心、神经网络加速核通过片上总线互联形成,峰值算力达到16TOPS(INT8)。FCOS模型是一个先进的单阶段无锚框目标检测深度学习模型,该模型首次提出的核心原理已经被一些新的目标检测网络模型采用。该文研究FCOS深度学习模型在该人工智能芯片上的部署,并研究片上存储器大小、DDR带宽、DDR配置、算力、数据类型等因素对FCOS深度学习模型部署的性能和检测效果的影响。可以为深度学习模型部署技术研究人员、人工智能芯片设计人员提供参考。

AI chip is a kind of chip that can perform artificial intelligence related computing efficiently.The 38th institute of CETC developed an AI chip targeted at deep learning inference tasks at edge devices,and its main application fields includes object detection in radar images,intelligent image processing in color separator devices,etc.The chip is a heterogeneous SOC chip that is composed of central processing unit and neural network accelerator,which are connected together by on-chip bus.Its peak performance achieves 16TOPS(INT8).FCOS is an up-to-date single stage and anchor free object detection deep learning model,the mechanisms firstly proposed by the model have been utilized in some newer object detection models.The application of FCOS on the chip is studied,and the influence of on-chip memory size,DDR bandwidth,DDR configuration,computing power,data type and other factors is thoroughly studied.This work will provide insights to researchers on deployment of deep learning models and also to designers of AI chips.