Factorial Approach to Assessment of GPU Computational Efficiency in Surrogate Models

Surrogate models are very useful to replace very accurate but time-consuming huge data numerical models. The process of construction and optimization of surrogate models may require large computational power. It may be delivered by multi-core CPU or by massively multi-core GPGPU. This paper presents a consistent approach to make quantitative assessment of the computational efficiency with different hardware configurations: with one and multi-core CPU and with GPGPU card. The design of experiments factorial approach is used to analysis of the obtained data. The linear main effects model with two-way interaction is identified. The results show that the investment to multi-core CPU and GPGPU cards simultaneously is impractical due to negligible effects of CPU efficiency which is masked by dominated GPGPU performance.