Acceleration of Quantum Transport Simulations with OpenCL

The Recursive Green’s Function (RGF) [1] is a subset of methods for calculation of the Non-Equilibrium Green’s Function (NEGF) [2], which is popularly used for simulations of quantum transport in nanoscale devices. The computation of RGF involves repeated multiplication of dense & complex matrix blocks as illustrated in Figure 1 (particularly Step 3), so enhancement of corresponding performance is critical to accelerate NEGF simulations. This work introduces acceleration techniques for RGF computations using GPUs, and validates the practicality of suggested techniques with solid sets of benchmark results. The code has developed with OpenCL [3] and CLBlast [4], using an in-house tight-binding simulation code package [5] as a baseline. Performance validation has been conducted with test cases summarized in Table 1. As Figure 2 shows, our code modernization that is conducted with OpenCL & CLBlast ZGEMM against NVIDIA & AMD GPU devices can speed-up the whole RGF computation and the hotspot up to 18.63 and 48.73 times, respectively, against the case where the computation has been done in host with hybrid parallelization based on MPI-OpenMP.