50% sparsity is almost certainly already being used given that it is accelerated in current nvidia hardware both at training time (usable dynamically through RigL, which also addresses initial conditions) and inference time.