Research On Computing Platform Of CPU/GPU/FPGA For Big Data

Big data is already related to people's lives,and it is closely related to science and technology,and with the Internet and the universal development of things,big data will bring greater commercial value to modern society.In the era of big data,due to the large amount of data processing and data processing algorithms to improve,only the use of CPU to analyze the massive data will lead to inefficiency.So people begin to build a new computing architecture to improve the performance of large data processing.Firstly,people use GPU in large data processing.Although the GPU does not have complex control logic as CPU,but it have a large number of computing resources,which CPU dosen't have.And GPU structure type is very uniform,suitable for large amounts of data in parallel computing.Subsequently,the FPGA began to excite people's attention.FPGA also has a wealth of computing resources,and has low power consumption,high throughput and so on.A cluster of multi-block FPGAs will have GPU-like image processing capabilities and lower power consumption.In this context,this paper uses FPGA and GPU two different types of chips,to make two kinds of computing platform with CPU,which are CPU-FPGA and CPU-GPU.In order to deal with the computational challenges of today's big data,we run floating-point matrix multiplication in the CPU-FPGA computing platform and run the most common machine learning algorithms of big data analysis in the CPU-GPU computing platform.In the CPU-FPGA computing platform,we solved the problem of CPU and multi-block FPGA communication,ARM and FPGA communication problems,and use Verilog hardware description language design and implementation of the floating-point matrix multiplication system.The entire computing platform can be achieved,the CPU send floating-point matrix which need to be calculated to FPGA by Ethernet.After calculated in FPGA,the results will be sent to CPU.In the CPU-GPU computing platform,we propose a parallel training method for neural network and convolution neural network algorithms in machine learning.According to our proposed parallel training method,we use the heterogeneous programming language OpenCL in the CPU-GPU computing platform to run.Both of these algorithms are used for handwritten numeral recognition.We compare the results of parallel training of neural network and convolution neural network with their serial training results respectively.The results show that the two machine learning algorithms running in the CPU-GPU computing platform can be used in almost no loss of performance Cases,reducing their training time and improving energy efficiency.