USING A GPU FOR REAL-TIME DIGITAL SIGNAL PROCESSING

Abstract

This paper is devoted to the development of energy-efficient implementations of digital signal processing
algorithms in MIMO radar for estimating target parameters on computers with different architectures.
In accordance with the global trend, the possibility of using computers with parallel architecture for
digital processing of broadband radar signals is being considered. The authors proposed an implementation
of the procedure for processing the reflected signal of MIMO radar using the technology of general computing
on graphics cards (GPGPU). The performance of the developed solution was assessed on various GPUs with different microarchitectures. A criterion for evaluating the performance of a processing algorithm is
proposed in the form of the ratio of the algorithm's throughput to the peak throughput of the computer's
memory. A numerical assessment of the efficiency of using the computer's memory bandwidth of the developed
algorithm was carried out in comparison with known implementations on the GPU. The purpose of this
work is to detect and evaluate target parameters in real time using MIMO radar, using a commercially
available computer with the minimum possible weight and size characteristics. To achieve the set research
purpose, the following problems were solved: – selection and adaptation of algorithms that allow the assessment
of target parameters in MIMO radar; – implementation of selected algorithms taking into account the
architecture of the computer, allowing for an assessment of the target background situation in real time; –
assessment of the performance of the resulting solution. In the process of developing an algorithm for digital
processing of a MIMO radar signal, several options for implementing the algorithm were analyzed taking
into account the architecture of a parallel computer, which made it possible to process a radio image frame
consisting of 8 million complex samples in less than 50 ms. by NVIDIA Jetson AGXXavier GPU. The inverse
relationship between frame processing time and the peak GPU memory bandwidth is shown. A criterion for
evaluating the performance of the processing algorithm is proposed. A numerical assessment of the efficiency
of using the computer's memory bandwidth of the developed algorithm was carried out in comparison with
known implementations on the GPU. The gain of the developed algorithm is on average 5 times compared to
the results obtained by other authors. Compared to an FPGA, implementing 2D FFT on a GPU is 17 times
faster. The practical significance of the functional software developed by the authors does not impose any
restrictions on the number of receiving and transmitting channels and can be used for signal processing in
MIMO radars with a large number of channels.