On the Effect of Using Multiple GPUs in Solving QAPs with CUDA