Accelerating density functional calculations with graphics processing unit

An algorithm is presented for graphics processing units (GPUs), which execute single-precision arithmetic much faster than commodity microprocessors (CPUs), to calculate the exchange-correlation term in ab initio density functional calculations. The algorithm was implemented and applied to two molecules, taxol and valinomycin. The errors in the total energies were about 10(-5) a.u., which is accurate enough for practical usage. If the exchange-correlation term is split into a simple analytic model potential and the correction to it, and only the latter is calculated with the GPU, the energy error is decreased by an order of magnitude. The resulting time to compute the exchange-correlation term is smaller than it is on the latest CPU by a factor of 10, indicating that a GPU running the proposed algorithm accelerates the density functional calculation considerably.