Deep neural network computing system and method

Abstract

This application discloses a deep neural network computing system and method. The system is composed of a CPU for receiving target data and carrying out processing by using a deep neural network to obtain input layer data of the deep neural network and an FPGA for performing a calculation link by using a GEMM calculation node and a parallel operation program in the deep neural network and calculating the input layer data to obtain a calculation result. According to the invention, the GEMM computing node of the deep neural network is transplanted to FPGA and the CPU receives the target data inputted by a user. On the basis of the deep neural network, the target data are transformed into the input layer data of the deep neural network and the transformed data are sent to the FPGA; and the FPGA performs the calculation link by using the GEMM calculation node and the parallel operation program in the deep neural network and calculates the input layer data to obtain the calculation result, thereby completing the calculation. On the basis of the hardware characteristics of the FPGA, the FPGA completes the calculation link, so that the operation energy consumption is reduced substantially and the operation cost is lowered.

Description

technical field [0001] The invention relates to the field of deep neural network accelerated computing, in particular to a deep neural network computing system and method. Background technique [0002] The development of GPU general-purpose computing technology has attracted a lot of attention in the industry. Facts have also proved that in some calculations such as floating-point operations and parallel computing, GPUs can provide dozens or even hundreds of times the performance of CPUs. The standards for GPU general computing currently include OpenCL (Open Computing Language, open computing language), CUDA (Compute Unified Device Architecture), and ATI STREAM. Among them, OpenCL is the first open and free standard for general-purpose parallel programming of heterogeneous systems. It is also a unified programming environment that facilitates software developers to write efficient and portable codes for high-performance computing servers, desktop computing systems, and handh...

AI Technical Summary

Problems solved by technology

[0003] At present, the acceleration method of the deep neural network usually adopts the GPU acceleration method. Its highly optimized computing library cudnn and high-performance GPU parallel processing architecture make the acceleration performance of the deep neural network on the GPU platform very superior, but its high power consumption The energy consumption ratio is very low, which is also a great disadvantage

Images