“A naive implementation of the Transformer attention mechanism requires materializing intermediate matrices of size N x N, leading to quadratic scaling of memory reads and writes to slower GPU HBM.”