Serving highly sparse Mixture-of-Experts (MoE) models is inefficient on current HBM-based GPUs and XPUs