Abstract: Here proposed is a mean field annealing approach to the performance driven circuit partitioning, in which the object is to minimize the total routing cost between cells, the constraints are timing and area, and the partitions have intrinsic topological relationships. In the algorithm, a permute matrix is used to map the problem to neural network,and the energy function including object item, area constrained item,and the timing constrained item is presented, and then iteration procedure is put into practice with the mean annealing equation. Normalization of neurons proves that one cell only can be assigned to one partition. The algorithm is programmed with Visual C++, and experimental result shows that it is an effective method. In addition, the human and machine combined partitioning conception is also discussed.