Strange behaviour of OOM-Error in tensorflow-GPU

Question

I\'m trying to create a network and train it on my GPU. But during training I encountered a very strange behaviour. If I use batch_size = 24 everything is fine.