Compute-Optimal Is Not Cluster-Optimal

# Summary Researchers developed MOSAIC, a framework showing that optimizing AI models for computational efficiency (FLOPs) doesn't necessarily optimize them for actual cluster costs measured in GPU-hours, due to factors like hardware throughput and system goodput. The study found that sparse mixture-of-experts models are particularly susceptible to this mismatch—they become more loss-efficient as sparsity increases, but this doesn't account for real-world deployment costs on actual hardware clusters.