Making Analytic Double-Hybrid DFT Gradients Practical: A High-Performance Multi-GPU RI Implementation
Year: 2026
Type: preprint
Abstract: Double-hybrid density functional theory (DHDFT) is the most accurate rung within the density-functional hierarchy, but its use for molecular structure and dynamics remains limited by the high computational cost of analytic nuclear gradients. Here, we present, the first GPU-accelerated implementation of analytic DHDFT gradients. The implementation, developed in the Extreme-scale Electronic Structure System (EXESS), employs resolution-of-the-identity (RI) techniques throughout the calculation and evaluates the complete relaxed gradient on GPUs, including the modified Z-vector response and the explicit nuclear derivative of the exchange–correlation potential. The distributed multi-GPU formulation is performanceportable across NVIDIA and AMD architectures, supports basis functions through g angular momentum, and enables calculations through quadruple-ζ basis quality. The implementation is validated against numerical derivatives and the established CPU implementation in ORCA for B2PLYP and revDSD-PBEP86D4. For sufficiently large calculations, EXESS sustains more than 50% of the theoretical peak double-precision floating-point performance and exhibits efficient strong scaling across multiple NVIDIA H200 GPUs and AMD MI250X GCDs. Compared with ORCA on a modern Sapphire Rapids CPU node, a single H200 GPU provides speedups ranging from approximately 10× to 675×. These results extend high-performance DHDFT from single-point energetics to efficient exploration of molecular potential-energy surfaces at the large basis-set quality required for accurate perturbative correlation.
Source: ChemRxiv
Authors: Ross Pure, Giuseppe M. J. Barca
Institutions: Australian National University, Monash University
Language: English
Subfield: Atomic and Molecular Physics, and Optics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7212136715