Skip to main content
NASA Advanced Supercomputing (NAS) Division World-class modeling, simulation, and supercomputing in support of the NASA mission

New GPU-Based Supercomputer Speeds Computations for NASA
Research Projects

This animation showing the surface contours of the friction coefficient experienced by the NASA High-Lift Common Research Model is an example of the type of modeling and simulation work being done on the agency’s new GPU-based Cabeus supercomputer. Experts in the NASA Advanced Supercomputing Division are developing an automated tool for the agency’s Launch, Ascent, and Vehicle Aerodynamics (LAVA) software that will reduce the amount of hands-on work necessary to generate a mesh around complex geometries and increase the prediction accuracy of simulations of unsteady flow phenomena at a reasonable cost. Animation credit: Victor Sousa, Emre Sozer, NASA/Ames

A new graphics processing unit (GPU)-based supercomputer has been added to the stable of NASA high-end computing resources operated by systems experts at the NASA Advanced Supercomputing (NAS) facility at the agency’s Ames Research Center. Named Cabeus after a lunar crater near the Moon’s south pole, the system is already providing NAS users with several petaflops of computational power and will increase the agency’s capability to run artificial intelligence and machine learning workloads that can benefit from GPU technology.

In the first 10 weeks of Cabeus’ operation, NASA engineers have taken full advantage of its capabilities, including use of the agency-developed FUN3D code for a wide range of GPU-based applications. These include wall-modeled large eddy simulations (WMLES) of high-lift flows for transport aircraft, launch abort analyses for the Artemis program, aeroelastic flutter applications, and simulations of ground test facilities.

Photo of the Cabeus supercomputer racks.
The new Cabeus supercomputer cluster adds 15 racks and 128 nodes to the main floor of the NASA Advanced Supercomputing facility at NASA’s Ames Research Center in Silicon Valley. Larry Tran, NASA/Ames 

“The computational power of these GPU resources allows users to perform extensive, high-resolution simulations and computations to assess predictive capabilities for a broad range of applications, with reduced computational time compared to traditional supercomputing resources,” said Eric Nielsen, lead for FUN3D development at NASA’s Langley Research Center in Hampton, Virginia.

Similarly, a GPU-based WMLES capability within the NAS-developed Launch, Ascent, and Vehicle Aerodynamics (LAVA) software framework “drastically improved the computational speed, and decreased the consumed power per simulation by several fold,” said Jared Duensing, acting chief of the NAS Division’s Computational Aerosciences Branch. The LAVA team’s work covers a wide variety of computational fluid dynamics problems, from modeling the complex flow fields that underlie turbofan engine aerodynamics and aeroacoustics to simulating the turbulent launch environment at NASA’s Kennedy Space Center.

The initial system, made available to users in late December 2023, contains 128 nodes utilizing AMD EPYC “Milan” processors and NVIDIA A100 GPUs. Each Cabeus node has four GPUs, for a total of 512 GPUs on the system. The current configuration adds five petaflops of performance to the NAS facility’s supercomputing capability.

Still, according to Duensing, significantly more GPU resources are required for multiple research teams to simultaneously tackle mission-critical, scale-resolving flow analyses for NASA projects. And others agree. According to the latest survey of scientists and engineers using the NASA high-end computing resources at the NAS facility, access to GPU-based systems is among their most important future requirements.

With that in mind, procurement of a next-generation GPU cluster is in the planning stages. In the short term, supercomputing systems experts at NAS plan to add two more AMD EPYC “Rome” based nodes with A100 GPUs, in addition to integrating 57 existing NVIDIA V100-enhanced Intel “Cascade Lake” and “Skylake” nodes—currently part of the Pleiades supercomputer—into Cabeus in Spring 2024.