On the Oak Ridge National Laboratory HPC system Summit (IBM Power System AC922, IBM POWER9 22C 3.07GHz, NVIDIA Volta GV100, Dual-rail Mellanox EDR Infiniband) the following scaling results have been achieved and published. For these results the ELPA release 2020.05.001 has been used. For the GPU calculations, all 6 V100 GPUs have been used, and each node was populated with 42 MPI-ranks.
Fig 4: Scaling of the 2-stage solver of the ELPA library for different real and complex, double-precision matrix sizes (100% of eigenvectors saught). The blue-solid line shows the computation of the CPU version on IBM Power9 processors. The blue-dashed line shows the execution of the ELPA2-GPU version with 6 Nvidia V100 GPUs per node. For comparison the ELPA1-GPU computations are shown in the red-dashed curves. The black-dotted lines show the theortical linear scaling.