Cost-efficiency of Large-scale Electronic Structure Simulations with Intel Xeon Phi Processors

Benefits of Intel Xeon Phi Knights Landing (KNL) systems in computing cost are studied with tight-binding simulations of large-scale electronic structures that involve sparse system matrices whose dimensions normally reach several tens of millions. Speed and energy usage of our in-house Schrodinger equation solver are benchmarked in KNL systems for realistic modelling tasks, and are discussed against the cost required by offload computing with P100 devices. Superiority in speed and energy-efficiency observed in KNL systems justify the practicality of bootable manycore processors that are adopted by nearly 30% of largest supercomputers in the world. With a demonstration of the strong scalability up to 2,500 nodes, this work serves as an useful case study that supports the utility of KNL systems for handling memory-bound applications including ours and other numerical problems that involve large-scale sparse matrix-vector multiplications, particularly compared to GPU-based systems.