Sparse matrices with 64-bit indices for addressing values, but 32-bit storage for column indices
E.g. PE has a matrix 471526400 x 471526400 with 3099301688 non-zeros. A 64-bit type is necessary to address the non-zero values and column indices in the global arrays, but we can store the column indices as 32-bit only. This would save about 25% of space (double + int vs double + long int).