Parallel execution of Lanczos algorithm in a CAM systolic ring

The author presents an efficient array processing scheme to exploit Lanczos algorithm for solving a large-scale eigenvalue problem. He shows that an array of processing elements with content addressable memories (CAM), connected in a simple ring topology, can solve the problem efficiently. The nonzero elements of a problem matrix are stored in a content addressable memory for an efficient storage and the required operations are executed in a pipeline fashion around the systolic ring. The algorithm consists of two steps: Lanczos triangularisation followed by QR method to obtain full eigenvalues and eigenvectors. It is shown that the speedup obtained in such a system is almost linear with the number of processing elements used when the problem is partitioned to be executed with relatively small number of processing element