Go to JKU Homepage
Implementation of signal processing algorithms on graphics cards
Michael Hofer
The detection of the presence of periodic components in measurement signals or the estimation of unknown parameters of a periodic signal is a signal processing task in many different areas, such as in speech signal processing, radar signal processing, in the production of endless material and metrology in general. There are various algorithms to evaluate such signals. The challenge with these algorithms is the very large processing time which typically prevents the real-time implementation or an implementation with an acceptable processing time. This is particularly relevant with the use of large sensor arrays, whose signals are to be analyzed in parallel.
Figure 1: Development of theoretical processing power of CPUs and GPUs
It should be noted that the above-mentioned algorithms are themselves highly parallelizable. With the use of parallel computing architecture of a graphics card (GPU) and its large number of available computing cores a considerable reduction of the total computation time can be achieved (Figure 1).
The goal was to estimate the performance of such GPUs for the solution of complex signal processing tasks in general as well as for the example of the periodic signals evaluation in measurement technology. A comparison of conventional, sequential implementations to parallel implementations in a specially designed hardware and software environment (CUDA) from the company NVIDIA shows the acceleration potential of the CUDA-enabled GPUs.
Keywords: detection, estimation, periodic signal, CUDA, graphics card, parallel implementation, NVIDIA, acceleration potential
The circulation of the project work is restricted from June 13, 2014 for a period of 5 years.
June 10, 2014
