Higher memory bandwidth enables the use of sparser Mixture of Experts (MOE) models, which reduces the required flops for a given intelligence level