A motion estimator uses many parallel Arithmetic-Logic-Unit (ALU) processors to simultaneously perform searches in many directions from a starting point. Each processor follows a different path outward from the starting point, generating sum-of-absolute differences (SADs) for each point in the path. A best SAD for the path is kept, along with an index into motion vector tables containing X,Y points for all paths. Current and best SAD's, thresholds, and indexes are stored in an ALU dedicated memory. When the number of best SAD's meeting thresholds exceeds a target, the current search-level ends. The index of the overall best SAD locates a new starting point, and a next-denser search-level is performed in the same manner, but over a smaller search area. Each processor calculates SAD's for one 16×16 macroblock, four 8×8 blocks, and 16 4×4 blocks and the net best SAD of these 3 types determines partitioning.