Abstract A high-performance multiprocessor is proposed that achieves very high speed by using a large number of elementary processing elements to permit efficient use of fine grain parallelism. Extremely high algorithm flexibility is maintained by means of a fast large reconfigurable generalized connection network constructed with an optical spatial light modulator. Static dataflow methodology permits high speed. Optical interconnections provide a bandwidth that is highly interference free to support fine granularity. An FFT and several forms of matrix-vector multiplier are implemented on the system with efficiencies approaching one. A conjugate gradient algorithm is modified to achieve maximum pipelining during an iteration. An implementation of this modified algorithm on the optical crossbar computer suggests that efficiencies of over 76% are feasible even when the problem dimension differs from the number of processing elements.