Introduction\n\nThe Gordon Bell Prize, awarded annually by the Association for Computing Machinery, has long been the gold standard for recognizing extraordinary achievements in high‑performance computing (HPC). Each year, a handful of research teams are selected to showcase how the latest advances in computer architecture, software, and algorithm design can push the boundaries of what is computationally possible. This year’s finalists, five in total, have taken the competition to new heights by harnessing NVIDIA‑powered supercomputers—namely Alps, JUPITER, and Perlmutter—to deliver breakthroughs that span climate science, fluid dynamics, and the broader movement toward open science. Their work exemplifies how the convergence of cutting‑edge hardware and collaborative research can accelerate discovery in ways that were unimaginable a decade ago.\n\nBeyond the sheer speed of the machines, the finalists’ projects demonstrate a commitment to transparency and reproducibility. By releasing simulation data, source code, and detailed methodological documentation, they are turning the Gordon Bell Prize into a catalyst for open‑science practices that benefit the entire scientific community. The result is a set of high‑impact studies that not only push the envelope of computational capability but also lower the barrier for other researchers to build upon their findings.\n\n## Alps: Climate Modeling at Scale\n\nAlps, a next‑generation supercomputer built around NVIDIA’s H100 Tensor Core GPUs, was the platform for one of the finalists’ most ambitious climate‑prediction models. The team leveraged the GPUs’ massive parallelism to run a global atmospheric simulation at a resolution of 0.25 degrees, a level of detail that was previously unattainable for full‑year runs. By integrating a sophisticated radiative transfer scheme and a high‑fidelity representation of cloud microphysics, the simulation produced temperature and precipitation forecasts that matched satellite observations with unprecedented accuracy.\n\nWhat sets this work apart is not merely the raw performance but the way the team optimized the code to exploit the GPUs’ tensor cores for matrix‑multiply‑accumulate operations that underpin the numerical weather prediction algorithms. The result was a 70 percent reduction in wall‑clock time compared to the same model on a traditional CPU‑based system, enabling researchers to run multiple ensemble members in parallel and thereby quantify forecast uncertainty more robustly.\n\n## JUPITER: Fluid Simulation for Aerospace\n\nJUPITER, another NVIDIA‑centric machine, was employed by a finalist team focused on fluid‑structure interaction in aerospace engineering. The project tackled the notoriously difficult problem of simulating the airflow around a hypersonic vehicle while simultaneously accounting for the deformation of the vehicle’s skin. By coupling a high‑order finite‑volume solver with a GPU‑accelerated structural dynamics engine, the team achieved a fully coupled simulation that ran in a fraction of the time required by legacy CPU codes.\n\nThe breakthrough came from a novel algorithm that partitions the computational domain into overlapping subdomains, each processed on a separate GPU. This approach eliminates the need for costly global communication and allows the solver to scale efficiently across thousands of GPUs. The resulting simulation revealed subtle aerodynamic phenomena—such as shock‑wave interaction with boundary layers—that were previously hidden in coarser models. These insights are now informing the design of next‑generation hypersonic aircraft, potentially reducing development costs and time to market.\n\n## Perlmutter: Multiscale Materials Science\n\nPerlmutter, a hybrid system that combines AMD EPYC CPUs with NVIDIA GPUs, served as the backbone for a finalist’s research into the behavior of advanced alloys at the atomic level. The team employed a hybrid quantum‑classical approach, where density‑functional theory calculations were offloaded to the GPUs while the surrounding molecular dynamics simulation ran on the CPUs. This hybridization allowed the researchers to simulate systems containing millions of atoms over nanosecond timescales—an order of magnitude larger than what was previously possible.\n\nThe scientific payoff is significant: the simulation uncovered a new mechanism of dislocation motion that explains why certain high‑entropy alloys exhibit exceptional strength and ductility. By publishing the simulation data and the open‑source code used to generate it, the team has provided a valuable resource for materials scientists worldwide, accelerating the discovery of next‑generation structural materials for aerospace and energy applications.\n\n## Open Science and the Gordon Bell Legacy\n\nA recurring theme across all five finalist projects is the deliberate embrace of open‑science principles. Each team has made their simulation outputs, source code, and detailed documentation available through public repositories and data portals. This openness serves multiple purposes: it allows independent verification of results, fosters collaboration across disciplines, and ensures that the computational methods can be adapted to new scientific questions.\n\nThe Gordon Bell Prize, traditionally focused on raw performance, has evolved into a platform that rewards not only speed but also the broader impact of research. By highlighting projects that prioritize reproducibility and data sharing, the prize is nudging the HPC community toward a more inclusive and collaborative future.\n\n## The Role of NVIDIA GPUs in HPC\n\nNVIDIA’s GPUs have become the workhorse of modern supercomputers, and the finalists’ achievements underscore why. The GPUs’ architecture, characterized by thousands of lightweight cores and high‑bandwidth memory, is ideally suited for the dense linear algebra operations that dominate scientific workloads. Moreover, NVIDIA’s software stack—CUDA, cuBLAS, cuFFT, and the emerging CUDA‑Fortran—provides developers with a rich ecosystem to accelerate code without sacrificing portability.\n\nBeyond raw hardware, NVIDIA’s commitment to open‑source initiatives, such as the RAPIDS suite for data science and the cuDF library for GPU‑accelerated data frames, has lowered the barrier to entry for researchers who may not have deep expertise in GPU programming. The finalists’ use of these tools demonstrates how a well‑integrated hardware‑software stack can transform a complex scientific problem into a tractable computational task.\n\n## Conclusion\n\nThe 2024 Gordon Bell finalists have set a new benchmark for what can be achieved when cutting‑edge hardware, sophisticated algorithms, and a culture of openness converge. Their work on climate modeling, aerospace fluid dynamics, and materials science not only pushes the limits of computational performance but also delivers tangible scientific insights that will shape policy, engineering, and technology for years to come. By making their data and code publicly available, they have turned the prize into a springboard for open‑science collaboration, ensuring that the benefits of their breakthroughs extend far beyond the walls of the supercomputing centers that hosted them.\n\n## Call to Action\n\nIf you are a researcher, engineer, or student interested in high‑performance computing, now is the time to dive into the world of GPU‑accelerated simulation. Explore the open repositories released by the finalists, experiment with the code on your own hardware, or contribute to the growing ecosystem of tools that make HPC more accessible. For institutions and funding agencies, consider supporting projects that prioritize both performance and openness, as this dual focus is the key to unlocking the next wave of scientific discovery. Together, we can harness the power of supercomputers to tackle the most pressing challenges of our time.