A new direct second-order Mø8ller-Plesset gradient algorithm is presented. By avoiding generation of molecular orbital integrals with three virtual indices, the memory requirement for a calculation with n basis functions and o occupied orbitals is reduced to on2, and overall computational savings are obtained. The algorithm has been implemented with distributed data on a massively parallel computer, the Intel Paragon. The parallelization scheme provides high parallel efficiency, and the performance of the algorithm is illustrated for calculations running on up to 64 processors.
Speedup curves have been obtained with the MP2 gradient program on the Intel Paragon.