Applications of GPUs for Signal Processing Algorithms: A Case Study on Design Choices for Cyber-Physical Systems

Abstract

Nowadays, signal processing algorithms are simulated generally using MATLAB. Their hardware implementation requires either application-specific IC (ASICs) or system on chip (SoCs). But there are severe constraints on producing such chips. Therefore, hardware implementation on graphical processing unit (GPUs) or field programmable gate array (FPGA) can provide the answer to the problem. Signal processing involves mathematical calculations in the form of algorithms, which are required to be implemented finally as stand-alone hardware to be used as a system. Recently, GPUs have increasingly being used for hardware implementation of signal processing algorithms. This is because they can be programmed easily with the help of open-source coding languages like Python, CUDA, or OpenCL providing cost benefits in terms of lower costs and generic programming. Also, they possess, in general, a greater number of cores as compared to ASICs or SoCs making GPUs multi-application platforms that can solve the problem of the lower yield factor of the ASICs. They are also better than ASICs and SoCs in terms of performance since it has a dedicated processor to handle 2D and 3D graphics, which comprises of polygons and polygonal transformations involving computationally dearer multiple floating-point operations. Hence, more complex signal analysis can be performed using them. Additionally, the massively parallel architecture of GPUs further enhances their high computing performance. There exist many GPU-accelerated applications that provide an easy way to high-performance computing (HPC). In light of the above discussion, this chapter intends to inform and help readers know properly about the hardware implementations of different signal processing algorithms, by showcasing appropriate hardware platforms and different open-source coding languages along with their implementation methodologies. Therefore, they will be able to appreciate the difference between hardware implementations using ASICs/SoCs or GPUs/FPGAs. As GPUs or FPGAs take a faster time-to-market approach, as no layout, masks, or other steps are required for the manufacturing, they possess a simpler design cycle and the most important, the feature of field reprogram ability.