Influences of SIMD architectures for scattered data interpolation algorithm

In this paper, we are investigating the performance of several IDW implementations on different SIMD architectures (Single Instruction Multiple Data). The SIMD architectures addressed in this paper are considered cost effective and readily available on the desktop as compared to super-computers such as the IBM Roadrunner or the Cray XT5. Two main classes of SIMD architectures can be identified from the range of products available now: (a) the architectures that are integrated in the CPU itself, such as SSE or AltiVec; and (b) the architectures located on a dedicated board such as the GPGPUs from Nvidia or ATI. This paper evaluates one SIMD architecture selected from each group. The evaluation is based on the performance measured on each architecture over various datatypes and different problem sizes.