摘要
人工智能技术在行业应用中大多依赖海量的训练数据和大规模服务器的算力支持。然而资源和能耗受限的在轨卫星难以满足当前人工智能模型动辄万亿级规模的存储和算力需求。因此,针对资源受限的星载计算,对人工智能设备和应用的高性能、轻量化需求越发凸显。从软件和硬件两方面来对神经网络进行设计及优化:软件层面进行模型量化压缩,缩减模型存储规模,并基于超长指令的编码格式设计和编译微指令;硬件层面基于Xilinx VC709平台实现了由微控制器与逻辑运算器组成的CNN加速器,加速器通过微指令控制来实现指令级并行,提升了系统灵活性,逻辑运算器通过DSP48E1计算资源构建带宽适配的脉动阵列实现运算级并行。
Most of the artificial intelligence technology in industry applications relies on massive training data and the computing power of large-scale server.However,in-orbit satellites with limited resources and energy consumption can hardly meet the storage and computing power requirements of the current trillion-scale artificial intelligence models.Therefore,for resource-constrained on-board computing,the demand for high performance and light weight of artificial intelligence devices and applications is becoming more and more prominent.In this paper,the neural network is designed and optimized from the aspects of software and hardware,at the software level,the model is quantized and compressed to reduce the model storage size,and the microinstruction coding format design and compilation based on Very Long Instruction Word are carried out.The hardware level is based on the Xilinx VC709 platform to implement a CNN accelerator composed of a Microcontroller Unit and a logic operator.The accelerator realizes instruction-level parallelism through micro-instruction control,which improves system flexibility.The logic operator uses DSP48E1 computing resources to build a bandwidth-adapted systolic array to achieve operation-level parallelism.
作者
尚金程
李浩东
徐瑞
相若彤
邓发俊
SHANG Jin-cheng;LI Hao-dong;XU Rui;XIANG Ruo-tong;DENG Fa-jun(Xi′an Aeronautics Computing Technique Research Institute,AVIC,Xi′an 710000,China)
出处
《航空计算技术》
2025年第6期103-108,共6页
Aeronautical Computing Technique
基金
国家重点研发计划重点专项资助(2025YFB3003400)。
