Definition
A way to compare neural architectures by holding the compute budget fixed and seeing which scales best.
An analysis methodology training models at multiple sizes under matched compute budgets, computing the compute-optimal frontier, and comparing architectures by their loss-versus-compute slope.