Administered by the National Science Foundation, the pilot has supported nearly 600 research projects, according to its website, and Sandra Gesing, executive director of the U.S. Research Software Engineers Association (US-RSE), said it’s looking to expand. Its original impetus came from an AI-focused executive order issued in October 2023 by former President Joe Biden, directing federal agencies to expand capacity for AI research, training and data sharing.
Several universities, 14 federal agencies and 28 industry partners all came together to support NAIRR with resources and expertise.
“It is really this mixture of funding agencies, industry, academia, national labs,” Gesing said. “I think that is a very unique situation and a real chance for collaboration.”
WHAT NAIRR OFFERS
For institutions and researchers without major high-performance computing infrastructure, NAIRR provides access to shared computing and data infrastructure as well as AI software and tools.
Before his NAIRR allocation, Liu said his group relied on a small number of local graphics processing units (GPUs) for large-model training.
“We have to test a lot of foundation models, and we have to do a lot of fine-tuning with those foundation models, using the data we collected from public domain and also data provided by our collaborator,” Liu said.
The process takes a lot of time and requires a lot of computation resources, he said. After receiving NAIRR support, the time to train the large language model (LLM) dropped from a few weeks down to a few days. According to Liu, that acceleration helped his group finish a paper on their epilepsy diagnosis model.
In a different domain, assistant professor Hongbo Zhang of Middle Tennessee State University (MTSU) worked with NAIRR on developing humanoid robots for agriculture. The project blends 3D simulation, custom data sets built from roughly 12,000 processed videos, and specific LLMs that translate visual and text inputs into actionable steps for a robot. Zhang said this requires greater computing power than MTSU has on campus, and NAIRR offered access to the Pittsburgh Supercomputing Center to meet the demand.
COLLABORATIVE OPPORTUNITIES
Gesing said NAIRR has a unique structure that allows partners to provide resources, not just use them. For example, the San Diego Supercomputer Center at the University of California, San Diego, provides access to GPUs, Amazon Web Services allows researchers to use its cloud platform, and Anthropic offers its models to research teams in need, according to the NAIRR resource website.
However, raising awareness of available resources remains a challenge. Gesing said researchers don’t always understand what is available to them or how those resources apply to their work.
“There are a lot of questions on how to start something out, because people don’t know which is the right resource for me to try out if they haven’t done it before,” she said.
At a session on NAIRR at the annual EDUCAUSE conference, Preston Smith, executive director of Purdue University’s advanced computing center, described two populations of NAIRR users: experienced users of high-performance computing and researchers new to AI.
“The needs of your computer science department and computer engineers doing that sort of stuff are different than the new domain scientist who is just learning how to apply AI,” he said.
For both groups, however, GPU scarcity is a limiting factor. Gesing said NAIRR portal operators have sought to address this by offering tiered GPU allocations and browser-based tools, as well as by hosting outreach events.
Looking forward, NAIRR hopes to establish an operations center that will sustain the program beyond the pilot phase, according to a solicitation issued by the organization in September.
