In the paper

We set the number of phases np to three and the number of nodes in each phase no to six.
Duringarchitecturesearch, we limit the number of filters (channels) in any node to 16 for each
one of the generated network architecture. We then train them on our training set using standard stochastic gradient descent (SGD) back-propagation algorithm and a cosine annealing learning rate schedule. Our initial learning rate is 0.025 and we train for 25 epochs, which takes about 9 minutes on a NVIDIA 1080Ti GPU implementation in PyTorch .

But in the code, the default parameter --n_nodes(number of nodes per phases) is four.
I set the channel to 16 and n_nodes to six ,but the search process is slow. So, I want to know if you have the concrete configuration about the 9 minutes.
Also , I find that the search code run slower in multiple GPU than run in a GPU,can you explain the phenomenon？
Tnank you very much!