Improving Energy Efficiency of Basic Linear Algebra Routines on Heterogeneous Systems with Multiple GPUs