Independent component analysis (ICA) is an important signal processing technique used to extract source signals from signal mixtures. Although useful in a wide range of problems, ICA is computationally expensive, and is therefore not suitable in many real-time or large data size applications. This paper presents a scalable parallel implementation of ICA in which computations are performed on graphics processing units (GPUs). An implementation using the programming toolkit OpenCL, as well as local memory and memory coalescing optimizations, increase ICA efficiency, and potentially improve its utility in data-intensive applications.