Both methods utilize the MUMPS solver for parallel matrix solving, matrix parallelism aims for full parallelization including assembly and finalization, whereas matrix_centralized keeps these phases sequential, which might be necessary for certain problem types that do not lend themselves well to domain partitioning, although potentially at the cost of higher memory usage.
For more details, please go through the knowledgebase documents at: https://simcompanion.hexagon.com/customers/s/article/Difference-between-the-matrix-and-matrix-centralized-parallel-processing-methods