Experiments with GPU-acceleration for solving a radiative transfer problem

Abstract

High performance computing systems are increasingly incorporating the computational power provided by accelerators, especially GPUs. With the programmability of GPUs greatly facilitated by OpenCL or NVIDIA's CUDA, with support for full double precision on GPUs, many challenging problems are benefiting from these processing units. It is well-known that memory latency is the speed limiting factor on GPUs. To hide memory latency, kernel instances must be executed in parallel on the same core, making sparse data more difficult to deal with than dense data. In this work we examine the numerical solution of a radiative transfer problem. We show that integral problem formulations relying on sparse linear algebra computations can benefit from the computing power of such devices, achieving an average speedup of 50x when compared to a representative CPU implementation. © 2014 Springer International Publishing.