GW Computations via GPU-GPU Message Passing: A path to High-throughput

The GW approximation uses a screened Coulomb interaction to calculate the self-energy of a system of interacting electrons. The resulting band structures and absorption spectra compare favorably with experiment, but the computations are currently too expensive to be used in high-throughput ab initio electronic structure calculations. In this talk we describe our implementation of the GW approximation, which uses GPU-GPU interprocess communication for acceleration, and discuss its advantages over traditional MPI code. In addition, we describe our "ATLAS"-style method of optimizing the code to prevent warp divergence on arbitrary CUDA-enabled GPU architectures, and end with an estimate of when this method will be cost-effective for high-throughput calculations assuming continued exponential growth in CUDA cores per GPU.