“A significant portion of the time spent in standard Transformer attention layers is on data movement (reading and writing to GPU memory), not on computation (FLOPs).”