Sept. 11, 2023 — As in previous years, the 2023 Gordon Bell Prize (GBP) drew a number of exciting submissions. From them, the Association for Computing Machinery (ACM) GBP Award Committee selected six finalists. Chester Gordon Bell (pictured above) is the namesake of the prestigious award that honors outstanding achievements in HPC.

The GBP is awarded to the most valuable scientific computation as demonstrated using state-of-the-art software and hardware technologies on world-leading supercomputers. The GBP represents various aspects of evaluation, such as the importance of target problems, performance optimization, maximum utilization of target system performance, and knowledge given for widely spread platforms.

This year, the Gordon Bell finalists’ work spans various applications, including materials science, fluid dynamics, nuclear simulation, seismic processing, and biomolecular simulations. The hardware platforms also include world-class systems: Frontier (ORNL, USA), new Sunway System (SC Wuxi, China), LUMI (EuroHPC/CSC, Finland), Leonardo (EuroHPC/Cineca, Italy), Cerebras CS-2 (KAUST, Saudi Arabia), and Perlmutter (NERSC, USA).

The following summarizes this year’s Gordon Bell finalists via brief descriptions of the work. Please note that the results or system sizes will not be finalized before the teams’ final submissions in August. The ACM Gordon Bell Prize award winner will be announced as part of a special ceremony at SC23 this November in Denver.

FINALIST 1

Large-scale Materials Modeling at Quantum Accuracy: Ab Initio Simulations of Quasicrystals and Interacting Extended Defects in Metallic Alloys

Sambit Das, Bikash Kanungo, Vishal Subramanian, and others (eight authors total) as part of a team that includes the University of Michigan, Indian Institute of Science, and Oak Ridge National Laboratory

In this work, the team developed a mixed method to combine density function theory (DFT) and the quantum many body (QMB) problem using a machine learning technique. The effort achieves high accuracy of calculation and affords large-scale modeling with the inverse-DFT that links the QMB method to DFT. They realized the ground-stage energy calculation while keeping the accuracy commensurate with QMB, using more than 60% of resources on the Frontier supercomputer housed within the Oak Ridge Leadership Computing Facility.

FINALIST 2

Towards Exascale Computation for Turbomachinery Flows

Yuhang Fu, Weiqi Shen Jiahuan Cui and others (20 authors total) as part of a team from Zhejiang University, Tsinghua University, National Supercomputing Center in Wuxi, Taiyuan University of Technology, Xi’an Jiaotong-Liverpool University, University of Cambridge, University of Florida, and University of Illinois Urbana-Champaign

The team developed a new large eddy simulation code to solve compressible flows in turbomachinery. They applied it to NASA’s grand challenge problem with a high-order unstructured solver for a high-pressure turbine cascade of 1.69 billion mesh elements and 865 billion degrees of freedom. The code has been calculated on Wuxi’s new Sunway supercomputer with extreme many-cores per node, up to 19.2 million cores, where each computation node consists of 384 calculation cores and six control cores.

FINALIST 3

Exascale Multiphysics Nuclear Reactor Simulations for Advanced Designs

Elia Merzaria, Steven Hamilton, Thomas Evans, and others (12 authors total) featuring a team from Pennsylvania State University, Oak Ridge National Laboratory, Argonne National Laboratory, and University of Illinois at Urbana-Champaign

The team simulated an advanced nuclear reactor system coupling radiation transport with heat and fluid simulation, including the high-fidelity, high-resolution Monte-Carlo code, Shift, and the computational fluid dynamics code, NekRS. Nek5000/RS was implemented on ORNL’s Frontier system and achieved 1 billion spectral elements and 350 billion degrees of freedom, while Shift achieved very high weak-scaling on 8192 system nodes. As a result, they calculated six reactions in 214,896 fuel pin regions below 1% statistical error, yielding first-of-a-kind resolution for a Monte Carlo transport application.

FINALIST 4

Exploring the Ultimate Regime of Turbulent Rayleigh–Bénard Convection Through Unprecedented Spectral-element Simulations

Niclas Jansson, Martin Karp, Adalberto Perez, and others (12 authors total), featuring a team from KTH Royal Institute of Technology, Friedrich-Alexander-Universitat, Max Planck Computing and Data Facility, and Technische Universität Ilmenau

The team developed a high-fidelity spectral-element code, Neko, which is essential for unprecedented large-scale direct numerical simulations of fully developed turbulence—all while maintaining high-performance portability on GPU-accelerated platforms. They applied a GPU-optimized preconditioner with task overlapping for the pressure-Poisson equation and in situ data compression. They also conducted initial runs of Rayleigh–Bénard convection at extreme scale on the LUMI and Leonardo supercomputers with up to 16,384 GPUs via a sophisticated workflow control.

FINALIST 5

Scaling the “Memory Wall” for Multi-dimensional Seismic Processing with Algebraic Compression on Cerebras CS-2 Systems

Hatem Ltaief, Yuxi Hong, Leighton Wilson, and others (six authors total) as part of a team from King Abdullah University of Science and Technology and Cerebras Systems Inc.

This work exploits the high-memory bandwidth of artificial intelligence (AI)-customized Cerebras CS-2 systems for seismic processing by leveraging the low-rank matrix approximation to fit the problem on SRAM (static random-access memory) wafer-scale hardware and use many wave-equation-based algorithms that rely on multidimensional convolution operators. As a result, the team implemented a standard seismic benchmark dataset into the small local memories of Cerebras processing elements, extrapolating a worst-case load-balanced whole application execution to 48 CS-2 systems on 35,784,000 processing elements. This is a significant example of applications run on AI-customized architectures that can enable a new generation of seismic algorithms.

FINALIST 6

Scaling the Leading Accuracy of Deep Equivariant Models to Biomolecular Simulations of Realistic Size

Albert Musaelian, Anders Johansson, Simon Batzner, and Boris Kozinsky as part of a team from the Harvard John A. Paulson School of Engineering and Applied Sciences

The group developed the Allegro architecture to bridge the accuracy-speed tradeoff of atomistic simulations and enable the description of dynamics in structures of unprecedented complexity at quantum fidelity. This is achieved through a combination of innovative model architecture, massive parallelization, and model implementations optimized for efficient GPU utilization. Allegro’s scalability is illustrated by a nanoseconds-long stable simulation of protein dynamics and up to 44-million atom structure of a complete, all-atom, explicitly solvated HIV capsid on the Perlmutter system at the National Energy Research Scientific Computing Center. They achieved strong scaling up to 100 million atoms.

---

Source: Taisuke Boku, SC23