For Mixture-of-Experts (MoE) models, increasing sparsity from 1-in-16 to 1-in-256 can significantly save on computation