How the United States Can Increase Access to Supercomputing

To increase access to HPC resources for more AI researchers, Congress should increase funding for supercomputing to $10 billion over the next 5 years. Specifically, Congress should increase total NSF funding in HPC infrastructure to at least $500 million per year to match the current demand for time on NSF’s HPC resources, which is more than three times greater than current supply. In addition, Congress should increase total DOE funding in HPC infrastructure to at least $1.5 billion per year to match current demand for access to DOE’s HPC resources, which is three times greater than what DOE is currently providing.

To determine where to allocate these funds, NSF should first measure how states are using HPC resources for AI research and then fund HPC systems in states that have low levels of HPC availability but whose institutions are conducting high levels of AI research. In states where HPC availability is high, federal investments in more HPC resources will not be the most effective way to close the national gap between HPC supply and demand because either institutions in the state already have funding for HPC-enabled AI research and are using it, or they do not have research funding, which means access to HPC is not the problem, but rather research funding is. Focusing on states with low HPC availability but high AI research potential will allow the government to address instances where the gap between HPC demand and supply is greatest. Additionally, DOE and NSF should diversify the portfolio of HPC resources they are making available to AI researchers, including by exploring cloud computing options.

Maximizing returns on investment in HPC will require careful resource management driven by an understanding of what system requirements AI researchers need and how existing grantees are using HPC systems. DOE and NSF should require those institutions that receive funding to adopt HPC auditing tools such as the XDMoD tool that reports on how optimally institutions are using HPC systems. NSF should also annually collect community requirements and publish roadmaps that allow it to better set HPC priorities and make more strategic decisions that reflect user requirements.

Maximizing returns on investment in AI research will require mechanisms to effectively translate basic AI research into products and services for the marketplace. To this end, NSF should foster more public-private partnerships by tripling the number of awards it grants through its Partnerships for Innovation aimed at accelerating the path to market for new technologies, from 50 to 150 grants. Further, as part of its recent initiative to create AI research institutes across the country, NSF should support proposals that are of regional importance and foster collaboration and partnerships between universities, local businesses, and state and local governments.

Ensuring all individuals have equal opportunity to succeed in becoming the next generation of AI researchers will mean increasing access to HPC for groups that are traditionally underrepresented in science and engineering. NSF and DOE should support partnerships that coordinate the sharing of computing resources with Minority-Serving Institutions (MSIs) that include Historically Black Colleges and Universities, Hispanic-Serving Institutions (HSIs), and Tribal Colleges and Universities (TCUs), as well as re-establishing targeted grants that fund HPC resources at MSIs. In addition, NSF and DOE should lower the barrier women face in gaining access to the supercomputing resources by replicating the Blue Waters project at the National Center for Supercomputing Applications that created an HPC allocations category open to researchers at U.S. academic institutions who are women.

Finally, creating a well-prepared HPC workforce will require all students with computer science backgrounds to have clear, structured pathways into the HPC workforce. To ensure students with terminal two-year computer science degrees or who transfer from community colleges can seamlessly move into upper-division coursework at four-year colleges without having to spend time duplicating technical fundamentals, NSF should provide funding for consortiums of two-year colleges and four-year colleges to work together in developing structured HPC curriculums.