EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

GPU-Accelerated Fock Matrix Computation with Efficient Reduction

Satoki Tsuji, Yasuaki Ito, Haruto Fujii, Nobuya Yokogawa, Kanta Suzuki, Koji Nakano, Victor Parque, Akihiko Kasagi

Applied Sciences · 25.04.2025 · 15 (9) · 4779

DOI: 10.3390/app15094779 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Applied Sciences, online erschienen am 25.04.2025. Verfasst von Satoki Tsuji und 7 weiteren Autor:innen. Die Arbeit wurde bisher 8-mal zitiert (Crossref) und nennt 46 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

In quantum chemistry, constructing the Fock matrix is essential to compute Coulomb interactions among atoms and electrons and, thus, to determine electron orbitals and densities. In the fundamental framework of quantum chemistry such as the Hartree–Fock method, the iterative computation of the Fock matrix is a dominant process, constituting a critical computational bottleneck. Although the Fock matrix computation has been accelerated by parallel processing using GPUs, the issue of performance degradation due to memory contention remains unresolved. This is due to frequent conflicts of atomic operations accessing the same memory addresses when multiple threads update the Fock matrix elements concurrently. To address this issue, we propose a parallel algorithm that efficiently and suitably distributes the atomic operations; and significantly reduces the memory contention by decomposing the Fock matrix into multiple replicas, allowing each GPU thread to contribute to different replicas. Experimental results using a relevant set/configuration of molecules on an NVIDIA A100 GPU show that our approach achieves up to a 3.75× speedup in Fock matrix computation compared to conventional high-contention approaches. Furthermore, our proposed method can also be readily combined with existing implementations that reduce the number of atomic operations, leading to a 1.98× improvement.

Zitieren

APA

Tsuji, S., Ito, Y., Fujii, H., Yokogawa, N., Suzuki, K., Nakano, K., Parque, V., & Kasagi, A. (2025). GPU-Accelerated Fock Matrix Computation with Efficient Reduction. Applied Sciences, 15(9), 4779. https://doi.org/10.3390/app15094779

Vancouver

Tsuji S, Ito Y, Fujii H, Yokogawa N, Suzuki K, Nakano K, et al. GPU-Accelerated Fock Matrix Computation with Efficient Reduction. Applied Sciences. 2025;15(9):4779. doi:10.3390/app15094779

BibTeX
@article{Tsuji2025,
  author = {Tsuji, Satoki and Ito, Yasuaki and Fujii, Haruto and Yokogawa, Nobuya and Suzuki, Kanta and Nakano, Koji and Parque, Victor and Kasagi, Akihiko},
  title = {GPU-Accelerated Fock Matrix Computation with Efficient Reduction},
  journal = {Applied Sciences},
  year = {2025},
  volume = {15},
  number = {9},
  pages = {4779},
  doi = {10.3390/app15094779}
}
Kostenlos starten →