Network size on GPU takes double what is reported in the paper

[Mohamed Ezz]

I'm using the network architecture U-Net, the paper mentions that a 6GB Titan X was used to train it, but on my GPU it takes around 12GB, it barely fits into my GPU. 

So the moment I started to tweak the network parameters (add a few layers), there's not enough memory. Also a theano implementation takes less than 6GB in GPU.

I'm wondering what could have gone wrong in my setting. Can it be float percision ?

[Mohamed Ezz]

Any ideas ?