Computer scientists in the Earth Systems Science Center (ESSC) at The University of Alabama in Huntsville (UAH) have coordinated with the IBM Research AI Hardware Center to acquire and install a collection of high-performance computing hardware to enable advanced AI-powered research. The new system will include IBM prototype Artificial Intelligence Unit (AIU) computer chips called IBM Spyre that are designed for AI, enabling more efficient processing of AI models, training, and other data-intensive tasks. IBM expects the computing cluster to be delivered and installed in December 2024 in Robert “Bud” Cramer Research Hall, also known as the National Space Science and Technology Center.
Dr. Robert Lindquist, UAH Vice President for Research and Economic Development, will represent UAH at the IBM Research AI Hardware Forum in New York on November 18, 2024. He will deliver a joint plenary talk with Priya Nagpurkar, IBM Vice President of Hybrid Cloud and AI Platform Research, to highlight how UAH plans to use the AIU cluster to accelerate AI use in research performed at UAH.
Researchers will primarily use the cluster of IBM Spyre chips to support ongoing efforts to develop downstream applications of AI foundation models that use scientific data. Foundation models are large AI models that are pre-trained on massive datasets and become a “foundation” for future analysis. The models can be tailored to specific scientific domains and fine-tuned for various applications, making them effective at reducing the time, effort, and cost required to analyze these datasets. As a result, the models offer scientists new, expedient means of using data in many ways, such as increasing precision in weather forecasts, enhancing climate models, or monitoring natural disasters.
UAH computer scientists Dr. Sujit Roy, Dr. Amy Lin, and Iksha Gurung have guided coordination efforts with IBM Research along with UAH Atmospheric and Earth Science professor Dr. Udaysankar Nair. For the past two years, they have worked with NASA’s Interagency Implementation and Advanced Concepts Team (IMPACT) to forge a partnership with IBM Research to develop foundation models for science, including models for geospatial Earth observations and weather and climate. Led by Dr. Roy, the foundation model development team within IMPACT worked closely with IBM researchers to build the models. Dr. Lin offered guidance on how to scale the models for high-performance computing, while Dr. Nair’s scientific perspective helped verify that model outputs were scientifically valid.
Dr. Roy predicts that the AIU chip installation will facilitate significant advances in on-campus foundation model research. “With the new hardware, we can accelerate our experimentation on foundation model architecture. The AIU chips provide more efficient computation for experiments and scaling compared to our currently available computing resources.” Dr. Roy notes that, “Designing, training, and testing a foundation model can take nearly a year, so the enhanced processing power of the AIU chips can help condense that timeline.”
UAH researchers anticipate that the computing system will be up and running soon after installation, and they look forward to testing its capabilities. “The level of success we can attain all boils down to how well the system can accelerate what we’re already doing,” says Gurung. “Ultimately we aim to build and deploy inference services on top of the models, allowing users to make real-time projections.”
IBM is also interested in how UAH will apply these new computational tools in analyzing Earth science data and how well the chips will perform in this capacity. Significant energy is required to power the chips that analyze dozens of terabytes of satellite data to build foundation models. Initial tests have shown that running the geospatial model on the IBM Spyre cluster will enable major efficiency gains and save an estimated 23 kW of power a year, the equivalent energy consumption of 20 U.S. homes and 85 tons of carbon emissions.
"The new cluster at UAH will enable researchers to tune and test not only new applications for these AI foundation models, but also the performance of the IBM Spyre AIU accelerators,” said Jeff Burns, director of the IBM Research AI Hardware Center. “We’re looking forward to seeing how the AIU chips can speed up AI workflows in this exciting field and ultimately lead to more efficient model deployment.”
In the future, the hardware may be used in other capacities at UAH. “We hope to be able to integrate this computing system into educational opportunities at UAH,” says Dr. Nair.
“Foundation model research is at the intersection of AI and education.”
