Towards a More Efficient Use of GPUs