by Pacific Northwest National Laboratory
A recent collaboration among researchers from HUN-REN Wigner Research Center for Physics in Hungary and the Department of Energy’s Pacific Northwest National Laboratory, along with industry collaborators SandboxAQ and NVIDIA, has achieved unprecedented speed and performance in efforts to model complex metal-containing molecules.
The collaboration resulted in 2.5 times the performance improvement over previous NVIDIA graphics processing unit (GPU) calculations and 80 times the acceleration compared to similar calculations using central processing unit (CPU) methods. The research study, recently published in the Journal of Chemical Theory and Computation, sets a new benchmark for electronic structure calculations.
Accelerating molecular modeling
The research team’s efforts have enabled unprecedented calculations for complex biochemical systems, which include transition metal metalloenzymes. Such metal-containing catalysts are crucial in numerous industrial and biological processes, playing an essential role in facilitating chemical reactions.
These powerhouses of energy conversion are vital for many industries, including medicine, energy and consumer products. They accelerate chemical reactions, lowering the energy required and making processes more efficient and sustainable. Understanding and optimizing these catalysts is essential for addressing global challenges, such as clean energy production and environmental sustainability.
“When you have one or more metals in your system, then you have a lot of electronic states that are very close in energy, but behave differently, and that’s why it’s really important to make sure that you describe them accurately,” said Sotiris Xantheas, a PNNL Lab Fellow, a co-author of the research study, and a chemical physicist who leads the Center for Scalable and Predictive methods for Excitation and Correlated phenomena (SPEC), as well as the Computational and Theoretical Chemistry Institute at PNNL.
Highly correlated quantum chemistry calculations
The recent advances have been made possible by bringing together academic and industry experts with expertise in the development of tensor network state algorithms and high performance computing, led by Örs Legeza, a co-PI of the SPEC project, and his group at the HUN-REN Wigner Research Center for Physics, in Hungary, working with SandboxAQ’s team of scientists, led by co-author Martin Ganahl, to perform quantum chemistry calculations on NVIDIA GPUs.
The diverse team contributions underscore the importance of collaboration and point to an exciting future for the field powered by GPUs. For instance, this work implemented the ab initio Density Matrix Renormalization Group method, which describes physical properties of large, complex electronic structures on all GPUs within a single node for the first time.
The goal of the research was to achieve efficient and accurate solutions to the many-body Schrödinger equation. These algorithms are crucial for understanding the electronic structures of molecules and materials and require computational power only available on a few computing systems worldwide.
The group’s collective expertise and shared resources have helped push the boundaries of quantum chemistry, allowing for rapid iteration and refinement in the study of highly correlated complex chemical systems. The project illustrates the potential of large-scale calculations to revolutionize how scientists approach challenging quantum chemistry problems.
“With advancing computational hardware and the extension to multi-GPU, multi-node architectures are expected to enable even more comprehensive calculations beyond the current capabilities,” said Legeza, who also holds an appointment as a research fellow at the Institute for Advanced Study at the Technical University of Munich.
“The ongoing collaboration aims to adopt large-scale GPU-accelerated calculations, further enhancing the efficiency and accuracy of quantum chemistry computations, utilizing even more recent hardware developments.”
As noted in the published study, chemists today largely rely on their intuition because rapid, highly accurate calculations are often unattainable. The ability to quickly iterate on different choices of large active spaces enables a more systematic search. Today’s GPU computing frameworks, combined with AI-guided physics and new methods for generating training data for large quantitative machine learning models, are expected to contribute to applications in energy, sustainability and health.
“The combination of NVIDIA’s state-of-the-art hardware with cutting-edge simulation techniques like tensor network algorithms for quantum chemistry has the potential to unlock an entirely new field of discovery,” said Ganahl.
More information: Andor Menczer et al, Parallel Implementation of the Density Matrix Renormalization Group Method Achieving a Quarter petaFLOPS Performance on a Single DGX-H100 GPU Node, Journal of Chemical Theory and Computation (2024). DOI: 10.1021/acs.jctc.4c00903
Provided by Pacific Northwest National Laboratory