A Performance Analysis of Leading Many-Core Technologies for Cellular Automata Execution

We extend the panorama of performance analyses of CUDA, OpenCL and SYCL for the execution of Cellular Automata. To this end, we apply the SciddicaT landslide model to a real event by considering two complex topographic surfaces of different granularity, thus resulting in two simulations of different computing loads. For each technology, we developed a global memory and two tiled implementations of SciddicaT by adopting the Nvidia nvcc compiler for CUDA, the Nvidia implementation of the OpenCL standard and the CUDA back-end of the Intel DPC++ compiler for SYCL. The experiments, performed on three Nvidia accelerators, point out from good to optimal performances of SYCL compared to CUDA according to the newer device’s architecture. The carried-out Roofline analysis evidences high cache effects, pointing out greater advantages of tiled implementations for older architectures.