Performance evaluation of NPB and SPEC CPU2006 on various SIMD extensions

Abstract

Nowadays, almost all the processors are integrated with SIMD extensions, with which significant speedup is obtained for the programs in multimedia and scientific computation. The length of SIMD vector register has been increasing all the time. For instance, the original length of SIMD extension components is 64-bit in MMX. It then rises to 128-bit in SSE and further 256-bit in AVX. The new Intel Many Integrated Core (MIC) architecture supports 512-bits SIMD. Though a higher speedup is theoretically possible as the vector length increases, more complex and efficient instructions are required to support the vectorization. We analyze the vectorization performance of NPB and SPEC CPU2006 with the increase of vector length and different SIMD instruction sets of SSE, AVX, and IMCI, based on which some advice are given for the vector length and instruction set design.