GB300 - AI Task Force
Developing Talent and Technology
Developing Talent and Technology
The NVIDIA DGX™ GB300 supercomputer supports the NPS-NVIDIA Cooperative Research and Development Agreement (CRADA) signed in 2024 (one of NPS' 40+ industry CRADAs). NVIDIA donated the DGX GB300 to the NPS Foundation, which installed it at NPS along with funding for research projects that require its high compute and memory capacity. Read the press release.
The deployment of the DGX GB300 expands the institution's ability to provide students and researchers with access to frontier computational capabilities, preparing them for positions of increased technical and leadership responsibility within complex research and enterprise environments.
"Each generation of computing capability has expanded what NPS students and researchers are able to accomplish," said NPS Chief Information Officer Trenton Hancock. "The NVIDIA DGX GB300 represents the next evolution of our computing infrastructure, enabling researchers to tackle increasingly complex operational problems using artificial intelligence, modeling and simulation, and machine learning at unprecedented speed and scale.”
Quick Facts
- This NVIDIA DGX GB300 supercomputer is the first to be deployed within the U.S. military.
- NVIDIA designated NPS as one of its first two NVIDIA AI Technology Centers (NVAITC) in the U.S.
- This DGX GB300 continues NPS’ unparalleled legacy of pioneering advanced computing. In 1960, NPS received the first Control Data Corporation (CDC) 1604 (Serial Number 1) designed by Seymour Cray and used transistors instead of vacuum tubes.
- DGX GB300 pairs 72 NVIDIA Blackwell Ultra GPUs with 36 NVIDIA Grace CPUs, allowing it to process large volumes of data.
- FP4 tensor core performance is 1,080 PFLOPS (dense) and 1,440 PFLOPS (sparse).
- Since being established in 1909, NPS has developed leaders and research solutions that enable maritime advantage.
NPS, NVIDIA, and Industry Collaborations
Three views of the DGX GB300 in its cabinet are shown. The cabinet houses three racks, where the left rack contains the compute components—graphic processor units (GPU) and central processing units (CPU). The center and right racks contain the support components. (U.S. Navy photos by Dan Linehan)
The DGX GB300 was delivered to NPS in 2026. (Photo courtesy of Naval Postgraduate School Foundation)
The compute rack of the DGX GB300 was first installed. (Photo courtesy of Naval Postgraduate School Foundation)
System connections and testing began after the compute rack and two support racks of the DGX GB300 were installed. (Photo courtesy of Naval Postgraduate School Foundation)
Installed in 1960, the CDC 1604 at NPS was one of the first computers to fully use transistors instead of vacuum tubes. It stored data on large magnetic tape drives. (Photo courtesy of NPS Dudley Knox Library)
A naval officer conducts Fleet-related computations at the terminal of the CDC 1604. (Photo courtesy of NPS Dudley Knox Library)
NPS, NVIDIA, and Industry Collaborations
The NPS-NVIDIA CRADA aligns NVIDIA technology with NPS’ mission to educate defense leaders. DGX GB300 is among the most advanced computing systems in the world. It provides world-leading accelerated processing capability necessary for complex AI, modeling and simulation, autonomy, and data-driven research.
Teams from Vertiv, DDN, VAST Data, and others whose contributions of infrastructure, storage, and technical expertise helped accelerate deployment of the capability. The relationships between these key partners enabled NPS to rapidly field an advanced computing capability through private-sector collaboration, demonstrating a new model for delivering technology, expertise, and innovation.
Expanding the Boundaries of Computation
The DGX GB300 enables research and analysis that was not possible before. With massive processing power and next-generation networking, it permits advanced research into areas such as cyber, weather forecasting, sensor fusion, and logistics at scales and speeds previously unavailable at NPS.
Managed by NPS, the DGX GB300 is a shared research capability. Access for secure experimentation is also provided through approved research collaborations involving NPS faculty, students, government organizations, operational commands, laboratories, and appropriate industry collaborations aligned with defense priorities.
Use Cases Include
- Development of AI applications to challenges in relevant fields ranging from weather modeling, oceanic and operations research, cybersecurity, disaster resilience, and response planning
- Large-scale foundation model training on operational datasets
- Fine-tuning and adaptation of LLMs for naval planning, intelligence, and operations
- Multi-agent AI training for Fleet tactics and operations research
- Synthetic data-driven model training to overcome classified, sparse, or denied datasets
- AI-assisted scenario training using self-play and Red & Blue team learning loops
- AI safety, robustness, and adversarial resilience training for contested environments
- Rapid experimentation and assessment of emerging AI capabilities prior to Fleet transition
Power and Cooling
Power and Cooling
The DGX GB300 is a self-contained, integrated computing platform that operates within NPS’ existing power and water infrastructure capacities. At peak use, the DGX GB300’s compute components draw roughly 135 (kilowatts) kW. Anticipated daily operations will average between 50-80% of peak capacity.
Its liquid cooling system flows a heat-transfer fluid in a closed-loop to cool the compute components. This fluid circulates through heat exchangers, where a second closed loop filled with cooling water is pumped to the roof to be cooled. These closed-loop liquid cooling systems transfer heat away from the compute components similar to how a car's radiator system removes heat from the engine. The support components are cooled by a bank of fans mounted at the back of the cabinet.
Technical Specifications
| Specification | Details |
|---|---|
| Supercomputer Model | NVIDIA DGX GB300 |
| GPUs | 72 NVIDIA Blackwell Ultra GPUs |
| CPUs | 36 NVIDIA Grace CPUs |
| Compute Trays | 18 |
| GPUs | CPUs Per Compute Tray | 4 NVIDIA Blackwell Ultra GPUs | 2 NVIDIA Grace CPUs |
| NVLink Switch Trays | 9 L1 NVIDIA NVLink Switches |
| High Bandwidth Memory Size | 20 TB |
| High Bandwidth Memory Bandwidth | 576 TB/s |
| Fast Memory | 37 TB |
| FP4 Tensor Core Performance, Dense | Sparse | 1,080 PFLOPS | 1,440 PFLOPS |
| FP8/FP6 Tensor Core Performance, Dense | Sparse | 360 PFLOPS | 720 PFLOPS |
Read More About AI at NPS
Asset Publisher
