On one node of an IBM Minsky (2 Power8 CPUs and 4 Nvidia P100 + NVLink) the following scaling results have been achieved and published. The results have been obtained with the ELPA version 2017.11.001.
Fig 3: Scaling of the 1-stage solver of the ELPA library for different real, double-precision matrix sizes (100% of eigenvectors sought). The red solid line shows the scaling in the ELPA “cpu-only” version on the IBM Power 8 processors. The blue lines show the solution with the GPU version of ELPA and utilizing 2 GPUs (dashed) and 4 (solid). For comparision older results on an Intel E5-2680 v2 system in the cpu-only (red dotted) and with 2 Nvidia K40m GPUs are shown.