Developing novel algorithms and foundations for discrete structures including AI, graph algorithms, and discrete optimization
The Oak Ridge National Laboratory (ORNL) Discrete Algorithms Group is at the forefront of developing advanced computing solutions to address some of the most urgent scientific challenges facing the Department of Energy (DOE) science mission. The group aims to scale and expand the functionality of existing computational algorithms to make them compatible with the novel architectures of upcoming supercomputers and to optimize them for fields of particular interest to DOE, including neuromorphic and quantum computing.
Notably, the group created the Low Rank Approximation with Constraints at Exascale (LORACX), an innovative and foundational AI algorithm for nonnegative matrix factorization. The group demonstrated LORACX’s scalability on up to 8,192 Frontier nodes when processing a 16.3 × 16.3 million element dense matrix in 3 seconds, achieving 0.67 exaflops in double precision with timing synchronizations and 0.76 exaflops without synchronizations. The group applied LORACX to a protein k-mer dataset comprising 20 million protein sequences with 390,000 4-mer features and mapped crystal nuclei formation and propagation on molybdenum disulfide, thereby enabling the identification of structural motifs and the tracking of their evolution to provide valuable insights into nucleation events and phase transition dynamics.
Members of the group also developed HyperNeuro — a distributed accelerated neuromorphic simulator that can simulate 33 billion neurons and 33 trillion synapses in 78 seconds on 64,896 graphic compute dies across 8,112 nodes of the Frontier supercomputer, making it the fastest simulator in the literature and about 600× faster than the state-of-the-art SuperNeuro simulator.
Additionally, the research group is pioneering privacy-preserving computing methods, particularly in federated learning and differential privacy. These advancements enable collaboration among researchers who use sensitive data while ensuring data security and privacy. This capability is especially critical for applications in healthcare and energy grid management as well as in broader scientific research.
What distinguishes this group is their emphasis on practical implementation at scale. They have achieved impressive results on leadership-class supercomputers such as Frontier and Perlmutter while focusing on energy efficiency. Their work includes developing energy-aware profiling tools and optimization techniques, demonstrating a strong commitment to sustainable computing practices.
The group’s future goals include continuing to bridge the gap between theoretical advancements and practical scientific applications by combining cutting-edge AI capabilities with strong privacy measures and energy efficiency. The group aims to responsibly apply massive computational power to tackle critical global challenges in climate science, materials development, and secure scientific collaboration.
The group’s collaborative projects involve partnerships between various national laboratories and universities, including collaborations with ORNL core universities such as Georgia Institute of Technology; Virginia Polytechnic Institute and State University; University of Tennessee, Knoxville; North Carolina State Universities; Duke University; and Vanderbilt University. Other collaborators include the University of Oregon and Wake Forest University; industry partners IBM, AMD, HP, GE, and NVIDIA; and other DOE national laboratories. These partnerships underpin the group’s collective effort to innovate and share knowledge in this vital field.
