Hybrid OpenMP-CUDA parallel implementation of a deterministic solver for ultrashort DG-MOSFETs

Abstract

The simulation of ultrashort two-dimensional double gate metal-oxide semiconductor field-effect transistors and similar semiconductor devices through a deterministic mesoscopic, hence accurate, model can be very useful for the industry: It can provide reference results for macroscopic solvers and properly describe weakly charged zones of the device. For the scope of this work, we use a Boltzmann–Schrödinger–Poisson model. Its drawback is being particularly costly from the computational point of view, and a purely sequential code may take weeks to simulate high voltages. In this article, we develop a hybrid parallel solver for a graphics processing unit (GPU)-based platform. In order to accelerate the simulations, the Boltzmann transport equations are solved on GPU using the CUDA programing model, while the Schrödinger–Poisson block is performed on multicore CPUs using OpenMP. We have adapted the costliest computing phases to the GPU in an efficient manner, achieving high performance and drastically reducing the simulation time. We give details about the parallel-design strategy and show the performance results.