Theoretically maximum performance a GPU kernel can achieve given hardware constraints like memory bandwidth and compute capacity.