The original paper includes convolutional layer support in their future work & next steps. But it's not a foregone conclusion that the same speedup will occur.