New Grace Hopper Nodes Double the Power of NASA's Cabeus Supercomputer
As part of the NASA Advanced Supercomputing (NAS) Division’s goal to provide NASA with the computing power needed to support the agency’s missions and programs, 350 NVIDIA Grace Hopper nodes were added to the Cabeus supercomputer—the agency’s largest graphics processing unit (GPU)-based computing cluster. The new hardware was released to users earlier this month, providing more than 13 petaflops of computing power and more than doubling Cabeus’ total peak capability.
Each node contains a Grace Hopper GH200 superchip, which combines a 72-core Grace CPU with a Hopper H100 GPU with a memory-coherent, high-bandwidth, low latency connection. This large, fast connection is well suited for traditional computational fluid dynamics simulations—such as predicting the impact of in-flight icing on aircraft—and artificial intelligence/machine learning workloads.
In early 2024, the Application Performance and Productivity team worked with agency application developers to test the portability of several agency numerical analysis codes to the ARM CPU architecture for a variety of input datasets on a single-node test system. Initial results showed these codes experienced an overall speedup of 1.5-2.2 times in application performance compared to existing Intel Milan-based augmented with NVIDIA A100 GPUs at the NAS facility, demonstrating that the new hardware uses less energy while running the same job. After these successes, NAS then moved forward with procuring the 350 Grace Hopper nodes, with 15 racks installed at the main NAS facility at NASA’s Ames Research Center in September 2024, before undergoing the long process of integration and benchmarking.
In addition to the new Grace Hopper hardware, NAS teams also migrated 57 existing Intel Cascade Lake- and Skylake-based NVIDIA V100 GPU-enhanced nodes from the Pleiades supercomputer to Cabeus, consolidating all the facility’s GPU-powered computing resources. With these two expansions, Cabeus now delivers 20.67 petaflops of peak computing power to researchers and scientists across the agency in supporting a wide range of science and engineering projects to meet NASA’s goals.
