Hacker News Reader

Top K sampling comes to mind, although it's nowhere nearly as important as matmult.

ranking models benefit from gpu impls of sort but yup they're not nearly as common/important as spmm/spmv