I'm going to go with not more than about 56MH/s ... based on memory bandwidth given the algorithm is memory bound.