AI Research Advocates Seek to Democratize US Supercomputing Infrastructure

Advocates for scaling up efforts to democratize access to AI research infrastructure aired their case during a House Science Committee hearing this week.

The Feb. 6 hearing focused on how federal science agencies can help researchers from less-resourced institutions tap into supercomputers needed for cutting-edge AI research.

The National Science Foundation launched a pilot program last month through which U.S.-based researchers can apply for time on such supercomputers, an effort known as the National AI Research Resource (NAIRR). It will also offer datasets, models, software training and user support.

Tess deBlanc-Knowles, special assistant to the director for AI at NSF, explained at the hearing that the pilot is being funded with existing agency resources and bolstered by in-kind contributions from a network of industry partners. Initially, the pilot will prioritize work to advance “safe, secure, and trustworthy AI,” as well as the application of AI to scientific challenges in health, the environment, and infrastructure sustainability, deBlanc-Knowles said.

The idea of creating a national AI research infrastructure that would increase access to computing power for all kinds of researchers gained traction through the National Artificial Intelligence Act of 2020, which established a task force to study how NAIRR should be structured. In its final report , published a year ago, the task force recommended that Congress commit $2.6 billion to the project over six years. The NAIRR pilot responds to an October 2023 executive order on AI that required NSF to launch the pilot within 90 days.

Committee members expressed general support for the NAIRR concept at the hearing, noting their desire for the U.S. to maintain a competitive global position.

“At the end of the day, our adversaries are not waiting for us,” Rep. Mike Collins (R-GA) warned the committee. “The Chinese Communist Party is implementing AI industrial policy on a national scale. It is Congress’s responsibility to ensure that America maintains its leadership role in AI, and that will require the academic community, industry partners, and small businesses to all work together to shape the future of this technology.”

“The committee looks forward to seeing how the NAIRR pilot at the National Science Foundation progresses over the next two years. With 11 federal agencies and 25 private sector and nonprofit organizations partnering in the pilot, it looks to be off to a good, promising start,” said Collins, who chairs the panel’s Research and Technology Subcommittee.

Given the increasingly high-cost of computing needed to train advanced AI models, Science Committee Chair Frank Lucas (R-OK) asked the witnesses how NAIRR will be able to keep up with the fast pace of development in the field and rapidly evolving industry standards for advanced computational power. Jack Clark, co-founder and head of policy at the AI company Anthropic, replied that NAIRR should allocate “a portion of its resources for a small number of big-ticket projects each year and adopt a consortium approach for picking what those are.” This would avoid a scenario where researchers are “unable to do their research due to running into computational limits,” he said.

Gina Tourassi, associate lab director for computing and computational sciences at Oak Ridge National Lab, backed the idea of picking a small number of large-scale projects via a consortium. Tourassi noted that this approach worked well for the public-private COVID-19 High Performance Computing Consortium, which helped to match urgent research projects with appropriate computational resources during the height of the pandemic.

Six computational resources are participating in the NAIRR pilot: Summit at Oak Ridge, Argonne National Lab’s Leadership Computing Facility, Frontera and Lonestar6 at the Texas Advanced Computing Center, Neocortex at the Pittsburgh Supercomputing Center, and Delta GPU at the National Center for Supercomputing Applications. However, Lucas pointed out that Summit is due to be decommissioned in October, less than halfway through the NAIRR pilot, and asked how NSF and DOE intend to plug this computation gap.

“That’s a very good question,” Tourassi replied, explaining that DOE “leaned in to donate, effectively, the supercomputer and keep its operations without additional funding” for several months — extending the life of the supercomputer beyond its slated decommissioning at the end of 2023. Depending on how NAIRR evolves and if there is sufficient interest and support, the department could “lean in once again,” Tourassi added.

Rep. Scott Franklin (R-FL) asked how the pilot will be evaluated and how Congress should determine whether the task force’s recommendation of $2.6 billion is an appropriate budget. DeBlanc-Knowles replied that key indicators of success will be the number of underserved researchers reached and the variety of projects they complete.

Chaouki Abdallah, executive vice president for research at Georgia Tech, urged Congress to make long-term spending commitments to AI research resources and training.

“Now is the time to realize that science and AI are just as critical to American security as are semiconductors and manufacturing,” Abdallah said, referencing the U.S. government’s $52-billion injection into the semiconductor sector through the CHIPS and Science Act.

Even at research-intensive R1 institutions such as Georgia Tech, which he said is spending around $200 million on AI-related research annually, the growing demand for computing resources is hard to keep up with. He noted how a single researcher at an AI institute at Georgia Tech exhausted the institute’s monthly budget of $1,000 for ChatGPT in just one day by switching from ChatGPT 3.5 to the more advanced ChatGPT 4.

The situation at R2 institutions is even more challenging, explained Louay Chamra, dean of the School of Engineering and Computer Science at Oakland University in Michigan. Chamra called for federal science agencies to create programs dedicated to supporting AI research at R2 institutions. “Furthermore, fostering collaboration between R-2 institutions and industry partners can bridge a gap between academic research and real-world applications, thus enriching the AI workforce,” he added.

Congressional approval will be needed to secure sustained funding for NAIRR. To that end, a bipartisan group of lawmakers in the House and Senate have introduced the Creating Resources for Every American To Experiment with Artificial Intelligence (CREATE AI) Act, which would formally establish NAIRR in statute. Leading sponsors of the bill appeared at the hearing and expressed hope that it will advance this year.

Committee Ranking Member Zoe Lofgren (D-CA), a sponsor of the bill, emphasized the importance of the U.S. government investing in AI research infrastructure and recognized that doing so would require significant federal investment. Lofgren warned the U.S. should not rely on a small handful of tech companies to realize the societal benefits of AI.

“Some of the private sector individuals, for example OpenAI, are skeptical that we can actually provide the computational power for non-big companies to do the research and compete. And I don’t think you have to be opposed to the private sector to believe, as I do, that this should not be owned by the biggest companies in the world,” she said. “This should be open to startups, this should be open to researchers, not just the big companies.”