Can the performance of GPGPU really beat CPU in evolutionary design task?

With the appearance of modern general purpose graphical processor units(GPU), a powerful and cheap architecture has entered the field ofscientific computation. This highly parallel architecture, formerlydesigned for floating point graphical operation acceleration, is nowbeing used for the acceleration of
various algorithms.

Duringthe past few years, various papers dealing with the utilization of GPUsin general purpose computing have been published. Even evolutionaryalgorithms have been accelerated [1, 3], among them genetic programmingand its variants. In order to achieve maximal performance of genomeevaluation, various approaches of candidate solution evaluation havebeen proposed. The genome can be evaluated as a program which can bedirectly downloaded into the GPU [1] or interpreted by using aninterpreter program running on the GPU [2]. Due to the architecturallimitations, the second method appears to be more promising incomparison with the previous one.

The GPUs are accessible viaspecial frameworks providing an interface between GPU and CPU. Thepurpose of these frameworks is to provide a comfortable programminginterface for rapid application development at different abstractionlevel. Thus, the utilized framework has a serious impact on theapplication's performance, since the higher abstraction the lowerperformance.

In this work [4] we focus on the acceleration ofCGP, which will be utilized for the evolutionary design of imagefilters. The application is written by using the nVidia CUDA framework,which allows a low-level access to the GPU resources. Several differentways, how to implement the candidate solution evaluation, with variousperformance impacts are discussed. Obtained results are compared with aCPU-based implementation. The experimental results show, that theaccelerated application does not exhibit the desired performance andeven in some cases is outperformed by a CPU-based application.