SIMD in, out and between execution timings

Three series of experiments were performed to measure the relative performance of SIMD and scalar data transfers. These experiments covered all 5 data types used in computation. Data was measured being moved between memory and register, and between registers. Data is relevant to Intel/AMD SIMD and SIMD register types xmm, ymm, and zmm.

The download document shows the programs used to obtain the measurements, the measured execution times obtained, and plots to indicate the relative execution times, and thus performance, of the different data transfers used. Floating point 32-bit and 64-bit data, together with 16-bit, 32-bit and 64-bit integer data are included.

This data is offered as evidence of the potential SIMD processing has relative to standard processing. This is a known. Now data is available to support this assumption.

Document version: Original -- November 2021