Optimising Convolutional Neural Networks Inference on Low-Powered GPUs