Partial least squares (PLS) methods have recently been used for many pattern recognition problems in computer vision. Here, PLS is primarily used as a supervised dimensionality reduction tool to obtain effective feature combinations for better learning. However, application of PLS to large datasets is hindered by its higher computational cost. We propose an approach to accelerate the classical PLS algorithm on graphical processors to obtain the same performance at a reduced cost. Although, PLS modeling is practically an offline training process, accelerating it helps large scale modeling. The proposed acceleration is shown to perform well and it yields upto ~30X speedup, It is applied on standard datasets in human detection and face recognition.