Design And Implementation Of Deep Learning Computing Platform Based On FPGAS

Deep learning,which is based on deep neural network,has become the most crucial technique in the field of artificial intelligence recently.In order to achieve significant accuracy improvement,the network is going deeper,while the model size of neural network keeps increasing.Large models not only mean intensive computing,but also lead to high energy consumption,which is a hard task for traditional general processors like CPUs.Recent works often explore specialized devices which are good at parallel computing.Due to its low power consumption,low latency,high performance and reconfigurability,FPGA has attracted more and more attentions in the field of neural network acceleration.This thesis is aimed at designing a deep learning computing platform based on FPGA to accelerate the computation of convolutional neural networks and cyclic neural networks.Firstly,by introducing the research background and basic theories of deep learning,the design idea of deep learning computing platform based on FPGAs are drawn.Next,the system architecture is introduced from the hardware and software level,and composition and function of the key modules of the system are briefly explained.Follows are the details of design and implementation of CNN and LSTM acceleration module,where some techniques are used for optimizing,such as fixed-point arithmetic,Winograd convolution and systolic arrays and so on.Finally,experiments are conducted to measure the whole system.This thesis integrated the CNN and LSTM acceleration engine into Caffe framework,forming a general deep learning computing platform for both CPU and FPGA to work together.The inference of neural networks can be accelerated on the platform without making complex modification to the pre-trained Caffe model.As the results show,the work conducted on FPGAs beats CPUs and GPUs both in performance and energy efficiency.