“Operations such as normalization, activations, and residual updates in Transformer training create a data movement bottleneck by repeatedly moving large intermediate tensors through global memory.”