Performance and energy consumption of a Gram–Schmidt process for vector orthogonalization on a processor integrated GPU

Thermal design power and vectorized instructions behavior

“…While all these studies compare the execution of applications without SIMD instructions and with a given generation of SIMD instructions, this paper provides a comparison of these instructions' generations (SSE, AVX, AVX2, and AVX512) over a representative range of HPC benchmarks. Moreover, many studies 5,6 explicitly start their experimental protocol by disabling turboboost capability as it leads to unexpected behavior. On the contrary, in this paper, we provide an in‐depth look into turboboost behavior for each generation of SIMD instructions over the aforementioned HPC benchmarks.…”

“…It means that the power consumption of supercomputers keeps increasing, resulting in larger and larger heat dissipation at the processor level. Consequently, because of thermal limits, it leads to a growing fraction b [1] a [1] b [2] a [2] b [3] a [3] b [4] a [4] b [5] a [5] + SIMD processor: several simultaneous operations FIGURE 1 Difference between SIMD and scalar operations. Example: addition of two vectors of dark silicon (2) in modern architectures.…”

“…Among the unexpected behaviors, this study provides actual examples illustrating that faster runs do not necessarily imply less energy consumption, and that using turboboost on a given architecture can save time and energy for different applications, while disabling turboboost consumes more time but less energy on another hardware architecture for the same applications. Moreover, this study exhibits experiments to deal with turboboost in depth, unlike other studies in the literature that start by deactivating the turboboost because of its impacts on performance and power consumption (5,6). All the data collected from the experiments of these study and presented hereafter are publicly available: https://gitlab.inria.fr/orgerie/greenavx-data.…”