Key points are not available for this paper at this time.
We present a general implementation of the resolution-of-the-identity (RI) and Cholesky decomposition (CD) representations of electron repulsion integrals within the coupled-cluster with single and double substitutions (CCSD) and equation-of-motion (EOM) family of methods. The CCSD and EOM-CCSD equations are rewritten to eliminate the storage of the largest four-index intermediates leading to a significant reduction in disk storage requirements, reduced I/O penalties, and, as a result, improved parallel performance. In CCSD, the number of rate-determining contractions is also reduced; however, in EOM the number of operations is increased because the transformed integrals, which are computed once in the canonical implementation, need to be reassembled at each Davidson iteration. Nevertheless, for large jobs the effect of the increased number of rate-determining contractions is surpassed by the significantly reduced memory and disk usage leading to a considerable speed-up. Overall, for medium-size examples, RI/CD CCSD calculations are approximately 40% faster compared with the canonical implementation, whereas timings of EOM calculations are reduced by a factor of two. More significant speed-ups are obtained in larger bases, i.e., more than a two-fold speed-up for CCSD and almost five-fold speed-up for EOM-EE-CCSD in cc-pVTZ. Even more considerable speedups (6-7-fold) are achieved by combining RI/CD with the frozen natural orbitals approach. The numeric accuracy of RI/CD approaches is benchmarked with an emphasis on energy differences. Errors in EOM excitation, ionization, and electron-attachment energies are less than 0.001 eV with typical RI bases and with a 10(-4) threshold in CD. Errors with 10(-2) and 10(-3) thresholds, which afford more significant computational savings, are less than 0.04 and 0.008 eV, respectively.
Epifanovsky et al. (Wed,) studied this question.