Аннотация:A new parallel algorithm for construction of a tensor of three-center electron repulsion integrals needed for the implementation of the resolution-of-the-identity Hartree-Fock method is proposed. The algorithm accounts for heterogeneity within a computing node, different evaluation times of matrix elements and their proper memory placement to ensure the optimal load balancing and therefore high parallel efficiency of the subsequent Fock matrix construction. Based on dynamic work scheduling and implemented in OpenMP, the algorithm was benchmarked on two AMD EPYC 7662 64-core CPUs employing up to 128 OpenMP threads for (H2O)_32 and (H2O)_48 water clusters; the 86x times speedup was demonstrated.