Toward electron-phonon physics at the exascale with a hybrid MPI-GPU-OpenMP framework

We report on the latest advances toward the exascale readiness of the EPW code, a Fortran/MPI program for first-principles studies of electron-phonon physics. Building on our GPU-accelerated implementation of the Wannier-Fourier interpolation of electron-phonon matrix elements, we have developed a hybrid framework that integrates MPI distribution, GPU offloading, and OpenMP multithreading to fully exploit heterogeneous supercomputing architectures. The new implementation delivers up to 29-fold speedups across Aurora (ALCF), Perlmutter (NERSC), and Vista (TACC) in terms of single-node performance compared to the CPU-only version of the interpolation using single-level MPI parallelism. Starting from a systematic analysis of the computational workload, we design an acceleration strategy that delivers efficiency, performance portability, and maintainability across NVIDIA, AMD, and Intel accelerators. We demonstrate large-scale calculations of electron self-energies and transport coefficients in strong-scaling tests with up to 6,144 GPUs on Aurora.