Two Hungarian researchers, Andor Menczer, a PhD student at ELTE, and Örs Legeza, a scientific advisor at the HUN-REN Wigner Research Centre for Physics, have set a new computational benchmark in the supercomputer simulation of complex quantum-physical systems. Their groundbreaking use of AI accelerators represents a significant milestone in quantum matter modeling, potentially streamlining experiments and industrial developments that have traditionally required substantial time and investment.
Photo: Adobe Stock
Their innovative simulation program achieves 250 quadrillion elementary operations per second, promising to reduce the cost of research in fields such as drug development and energy efficiency. Using the tensor network algorithm, Andor Menczer and Örs Legeza reached nearly a quarter of a PetaFlop on a single computer, with their findings published recently in collaboration with the US Pacific Northwest National Laboratory and startup industry partners NVIDIA and SandboxAQ. "This achievement with AI accelerators sets a new standard in computational quantum matter modeling, challenging the performance balance between classical and quantum computers," said Legeza.
The team reached 246 TeraFlops on the NVIDIA DGX-H100, comparable to the combined power of 80 high-performance, 128-core computers, or 700-1000 modern laptops. This performance is about half of the 0.6 PetaFlops capacity of Komondor, a domestic AI-enabled supercomputer. This breakthrough highlights the potential of new hardware tools for algorithms beyond those solely based on AI. (A joint press release on this accomplishment was issued by the US Department of Energy, Pacific Northwest National Laboratory, NVIDIA, and SandboxAQ.)
Performance could be further boosted by linking multiple computers together; with a multinode setup, reaching multi-PetaFlop scales becomes feasible. For context, in 2015, one of the world’s leading Japanese supercomputers achieved 10 PetaFlops. "The synergy of advanced mathematical algorithms and rapid IT advances is making it possible to study complex quantum systems that researchers once only dreamed of," says Legeza.
Beyond computational breakthroughs, the research also produced unprecedented precision results for biochemical systems with transition metal metalloenzymes. Metal-containing catalysts play a crucial role in many industrial and biological processes, driving essential reactions. These “powerhouses” of energy conversion are fundamental in industries spanning from medicine to energy generation. By accelerating chemical reactions, catalysts make processes more efficient and sustainable. "Optimizing these reactions is vital for tackling today’s global challenges, from green energy production to environmental sustainability," adds Legeza.
The new simulation approach is generating industrial interest, as the combination of the tensor network algorithm with AI-based methods creates a revolutionary environment for the pharmaceutical and chemical sectors. The substantial performance gains allow calculations that once took months to be completed in a day, providing a transformative toolkit for quantum chemical modeling.
In collaboration with engineers at NVIDIA and AMD, the team continues to optimize the algorithm on newer hardware, some of which is yet to be publicly available. "This joint achievement highlights the remarkable potential of industry-academia synergy," Legeza concludes.
